1
00:00:00,000 --> 00:00:10,040
Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the

2
00:00:10,040 --> 00:00:15,760
latest trends, tips and tools in the world of marketing AI, helping you get the best

3
00:00:15,760 --> 00:00:18,600
results from your marketing efforts.

4
00:00:18,600 --> 00:00:24,440
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:24,440 --> 00:00:25,800
Hello everyone.

6
00:00:25,800 --> 00:00:31,120
Welcome to Artificially Intelligent Marketing, a podcast where we talk about all things AI

7
00:00:31,120 --> 00:00:35,680
and marketing for marketers across the world so that they can leverage AI tools to get

8
00:00:35,680 --> 00:00:38,300
the most out of their work.

9
00:00:38,300 --> 00:00:44,120
This is episode eight and I'm back today with my very good friend Martin Broadhurst.

10
00:00:44,120 --> 00:00:46,680
Martin, I have missed your face.

11
00:00:46,680 --> 00:00:48,160
Well, here it is.

12
00:00:48,160 --> 00:00:50,080
Take it in, admire it in all its glory.

13
00:00:50,080 --> 00:00:52,800
It's been a long time, two whole weeks.

14
00:00:52,800 --> 00:00:56,240
That was quite the diary clash for both of us, wasn't it?

15
00:00:56,240 --> 00:00:58,240
It was a bit, but we're back.

16
00:00:58,240 --> 00:01:06,440
So the inane banter that hopefully at least five to 10% of listeners enjoy is back today.

17
00:01:06,440 --> 00:01:10,000
But also I've missed your AI insights, Martin, so I'm going to be glad to hear those too.

18
00:01:10,000 --> 00:01:12,440
Before we get into any of that, I do want to hear about the Caribbean.

19
00:01:12,440 --> 00:01:13,440
How was it?

20
00:01:13,440 --> 00:01:14,440
Yeah, beautiful.

21
00:01:14,440 --> 00:01:21,800
I don't know if you've ever been, but sitting on a beach, drinking a pina colada, reading

22
00:01:21,800 --> 00:01:25,560
the latest news on AI is the best way to read the latest news on AI.

23
00:01:25,560 --> 00:01:30,560
And I have to admit, I did listen to some Lex Freeman podcasts on the beach.

24
00:01:30,560 --> 00:01:36,520
So I was still nerding out on all this stuff, but I was doing it in the sun and I was doing

25
00:01:36,520 --> 00:01:37,520
it with a nice drink in my hand.

26
00:01:37,520 --> 00:01:42,440
Well, that sounds lovely because I was doing it in the British drizzle.

27
00:01:42,440 --> 00:01:45,280
Bit of drizzle.

28
00:01:45,280 --> 00:01:46,280
Yeah.

29
00:01:46,280 --> 00:01:47,280
Wow.

30
00:01:47,280 --> 00:01:48,760
I'm sorry to hear that.

31
00:01:48,760 --> 00:01:50,920
And we won't even talk about Derby County this week.

32
00:01:50,920 --> 00:01:53,280
Outside the playoffs.

33
00:01:53,280 --> 00:01:55,480
It's an absolute travesty.

34
00:01:55,480 --> 00:02:00,960
We absolutely won't talk about it because it would just make me weep.

35
00:02:00,960 --> 00:02:02,560
We can't have you weeping.

36
00:02:02,560 --> 00:02:04,320
Let's talk about what we're going to talk about.

37
00:02:04,320 --> 00:02:09,040
So this week's stories that we're going to cover, we are going to look at Stability AI

38
00:02:09,040 --> 00:02:15,120
launching their new open source large language model to rival the chat GPTs and GPT-4s of

39
00:02:15,120 --> 00:02:16,120
the world.

40
00:02:16,120 --> 00:02:19,600
Google announces its project Magi.

41
00:02:19,600 --> 00:02:25,360
What does this mean?

42
00:02:25,360 --> 00:02:28,560
It means changes to Google search are coming and we'll talk about that.

43
00:02:28,560 --> 00:02:34,160
Adobe has teased its AI driven video editing tools powered by natural language instructions,

44
00:02:34,160 --> 00:02:37,760
which the demo video for that looks rather cool.

45
00:02:37,760 --> 00:02:39,640
So we can talk about that.

46
00:02:39,640 --> 00:02:45,440
There was an economist article this week proposing an international agency to oversee AI and we've

47
00:02:45,440 --> 00:02:51,080
previously talked on the podcast about the challenges of trying to slow AI development

48
00:02:51,080 --> 00:02:57,080
down when there's so many different competing issues around companies competing with each

49
00:02:57,080 --> 00:03:01,000
other, geographical, you know, geographical political issues there as well.

50
00:03:01,000 --> 00:03:03,680
That's going to be an interesting one.

51
00:03:03,680 --> 00:03:08,440
Atlassian has infused AI into all of its products.

52
00:03:08,440 --> 00:03:14,280
So if you're a Jira user or a Confluence user, you can now tap into AI powered tools, which

53
00:03:14,280 --> 00:03:15,760
we'll talk about.

54
00:03:15,760 --> 00:03:21,520
And then we'll also talk a little bit about Snapchat's AI chatbot, which has perhaps gone

55
00:03:21,520 --> 00:03:26,360
as viral and if not more viral than some of the other chatbot style tools that have appeared

56
00:03:26,360 --> 00:03:29,820
since chat GPT came out at the end of last year.

57
00:03:29,820 --> 00:03:34,880
We are also going to have a lovely tool of the week this week where we're going to look

58
00:03:34,880 --> 00:03:41,000
at Anthropics Claude model and go into a bit more detail on that.

59
00:03:41,000 --> 00:03:43,240
Right let's get stuck in.

60
00:03:43,240 --> 00:03:48,880
Let's look first of all then at Stability AI launching its open source LLM model called

61
00:03:48,880 --> 00:03:53,860
Stable LLM to compete with the open AIs and Anthropics of the world.

62
00:03:53,860 --> 00:03:58,560
So this is seeing Stability AI get into the large language model game.

63
00:03:58,560 --> 00:04:01,920
If the name is familiar, it probably should be because they've already released quite

64
00:04:01,920 --> 00:04:06,520
a few well known models for different uses, perhaps the most famous of which is Stable

65
00:04:06,520 --> 00:04:12,520
Diffusion, which you can use for image generation.

66
00:04:12,520 --> 00:04:17,000
Martin and I were talking about this a little bit before we got on the recording here in

67
00:04:17,000 --> 00:04:25,560
terms of what it means to have open source models versus more closed models like GPT4.

68
00:04:25,560 --> 00:04:30,440
We'll dive into that in a bit more detail later, but with Stable Diffusion you can access

69
00:04:30,440 --> 00:04:35,880
the underlying model, remove some of the guard rails from it and do some cool stuff with

70
00:04:35,880 --> 00:04:36,880
it that you want to do.

71
00:04:36,880 --> 00:04:41,160
And so I guess the inference here is that that's going to be possible with Stable LLM.

72
00:04:41,160 --> 00:04:46,000
However, as TechCrunch said this week, the number of open source text generating models

73
00:04:46,000 --> 00:04:51,600
grows practically by the day as companies large and small vie for visibility in the

74
00:04:51,600 --> 00:04:54,920
increasingly lucrative generative AI space.

75
00:04:54,920 --> 00:04:57,120
So this is why is this important?

76
00:04:57,120 --> 00:05:00,780
It's kind of like, oh, and another one.

77
00:05:00,780 --> 00:05:06,600
But I think some quotes in the blog post from Stability AI are a little bit telling where

78
00:05:06,600 --> 00:05:11,720
they say, we open source our models to promote transparency and foster trust.

79
00:05:11,720 --> 00:05:15,960
Researchers can look under the hood to verify performance, work on interpretability techniques,

80
00:05:15,960 --> 00:05:19,140
identify potential risks and help develop safeguards.

81
00:05:19,140 --> 00:05:23,780
They also say open fine grain access to our models allows the broad research and academic

82
00:05:23,780 --> 00:05:30,360
community to develop these techniques beyond what is possible with closed models.

83
00:05:30,360 --> 00:05:38,200
So perhaps they're might a little dig in the ribs to the OpenAI team, whose name is OpenAI,

84
00:05:38,200 --> 00:05:45,800
but have been criticized recently, especially with the release of GPT-4 for how little information

85
00:05:45,800 --> 00:05:49,120
they've shared about how the model was built and trained.

86
00:05:49,120 --> 00:05:53,360
I think they've given some very good reasons in terms of, I know there's commercial reasons

87
00:05:53,360 --> 00:05:59,060
for that, but I think there's also some potential safety reasons for that.

88
00:05:59,060 --> 00:06:04,480
And it does feel a little bit like Stability AI are saying, hey, this is what OpenAI actually

89
00:06:04,480 --> 00:06:05,480
looks like OpenAI.

90
00:06:05,480 --> 00:06:08,480
What are your thoughts?

91
00:06:08,480 --> 00:06:10,640
Yeah.

92
00:06:10,640 --> 00:06:12,160
That's their whole model is that open source.

93
00:06:12,160 --> 00:06:13,560
There's a real commitment to it.

94
00:06:13,560 --> 00:06:20,440
You see actually a lot of engagement from the Stability AI research team and all the

95
00:06:20,440 --> 00:06:27,400
developers on Hugging Face, which is committed to open source AI ML projects.

96
00:06:27,400 --> 00:06:30,680
So yeah, it's nice to see that there is a company committing to it.

97
00:06:30,680 --> 00:06:38,660
I think the open source model for LLMs, that space is going to get quite interesting because

98
00:06:38,660 --> 00:06:46,480
you've also got Bloom as a large language model on the horizon.

99
00:06:46,480 --> 00:06:49,400
That's an open source model as well.

100
00:06:49,400 --> 00:06:53,360
So yeah, it's exciting to see what we can do with this.

101
00:06:53,360 --> 00:06:59,280
I think the ability for researchers to get under the hood, but also commercial operators

102
00:06:59,280 --> 00:07:07,720
to start shaping these models in the way that they want is also going to be quite interesting.

103
00:07:07,720 --> 00:07:16,040
That said, how many people are going to do that when you've got systems like GPT-4, which

104
00:07:16,040 --> 00:07:18,480
is so incredibly capable?

105
00:07:18,480 --> 00:07:25,000
It's a bigger, more capable model and is likely to prove to be a bigger, more capable model

106
00:07:25,000 --> 00:07:29,760
than the top level Stability LLM when that comes out.

107
00:07:29,760 --> 00:07:34,120
I think it's a hundred and seventy five hundred eighty billion parameter model that they're

108
00:07:34,120 --> 00:07:37,840
talking about releasing down the line.

109
00:07:37,840 --> 00:07:43,160
So we know that GPT-4 is bigger than that.

110
00:07:43,160 --> 00:07:49,880
When OpenAI is making customization of GPT-4, I was going to say so easy, that's definitely

111
00:07:49,880 --> 00:07:56,680
not the case, but there is vector embeddings, there is fine tuning available with that through

112
00:07:56,680 --> 00:07:59,680
the API.

113
00:07:59,680 --> 00:08:06,240
How many organizations just need that capability rather than having a model that actually can

114
00:08:06,240 --> 00:08:13,360
pull to pieces in an open source way, say pull to pieces, that's probably wildly exaggerating

115
00:08:13,360 --> 00:08:14,360
it.

116
00:08:14,360 --> 00:08:16,360
But you understand the point.

117
00:08:16,360 --> 00:08:18,400
Yeah, I get what you're saying.

118
00:08:18,400 --> 00:08:22,280
I think that leads on to the question then why is this important for marketers and what

119
00:08:22,280 --> 00:08:23,480
do they need to know?

120
00:08:23,480 --> 00:08:29,880
And fundamentally, open source LLM, large language models are more customizable because

121
00:08:29,880 --> 00:08:30,880
they're open source.

122
00:08:30,880 --> 00:08:35,920
And so that for developers of products will give them the capacity to potentially build

123
00:08:35,920 --> 00:08:39,520
on top of this and customize the model in different ways.

124
00:08:39,520 --> 00:08:45,240
I guess for companies with the necessary resources, they may even consider that they want to create

125
00:08:45,240 --> 00:08:49,040
an LLM of their own off the back of something like this.

126
00:08:49,040 --> 00:08:54,400
And to be fair, some of the other open source models like Llama and Alpaca and GPT-4 all

127
00:08:54,400 --> 00:09:00,120
and so stability AI is maybe a little late to the game here.

128
00:09:00,120 --> 00:09:04,920
But all of those are options for people who want to train their own model on their own

129
00:09:04,920 --> 00:09:06,080
data.

130
00:09:06,080 --> 00:09:12,560
The other thing with these open source models is you can run them on your computer, albeit

131
00:09:12,560 --> 00:09:20,760
in most cases very slowly, which might make it easier for people to use these models purely

132
00:09:20,760 --> 00:09:24,200
by not in effect having to use cloud-based resources.

133
00:09:24,200 --> 00:09:29,400
And I think even for some organizations, they're prohibited from using cloud-based LLMs like

134
00:09:29,400 --> 00:09:32,720
ChatGPT because you have to give data over to another organization.

135
00:09:32,720 --> 00:09:37,800
Whereas if you can run an open source LLM on your own computer and you're effectively

136
00:09:37,800 --> 00:09:43,960
running a local chatbot style version, you can obviously be much more confident that

137
00:09:43,960 --> 00:09:48,440
the information is safely stored on your computer.

138
00:09:48,440 --> 00:09:53,400
But yeah, what about these models without guardrails as well?

139
00:09:53,400 --> 00:09:57,980
We were talking a bit about that, Martin, because obviously there has been some criticism

140
00:09:57,980 --> 00:10:05,080
of ChatGPT and GPT-4 in terms of when it will refuse to do things or if it will say, oh,

141
00:10:05,080 --> 00:10:09,000
you know, just before we get into this, I'm not a doctor, I'm not a lawyer and all these

142
00:10:09,000 --> 00:10:12,840
other caveats that it slips in, which I completely understand.

143
00:10:12,840 --> 00:10:18,080
But I do think there are people out there who have the feeling that as responsible adults,

144
00:10:18,080 --> 00:10:22,120
they want to be trusted to interact with the tool and use their own good judgment, whether

145
00:10:22,120 --> 00:10:26,060
that is good or not, I'm not really in a position to say.

146
00:10:26,060 --> 00:10:31,520
But you were talking about some of how some of these models have been used without guardrails

147
00:10:31,520 --> 00:10:34,680
and how open sourcing allows you to explore the fringes.

148
00:10:34,680 --> 00:10:36,000
Yeah.

149
00:10:36,000 --> 00:10:42,680
So you can see this with the other stable or stability AI model that many of us will

150
00:10:42,680 --> 00:10:45,440
be more familiar with, stable diffusion.

151
00:10:45,440 --> 00:10:51,960
Stable diffusion, the text to image, image generator has been open sourced and the open

152
00:10:51,960 --> 00:10:58,800
source community does like to hack things to pieces and create things and build upon

153
00:10:58,800 --> 00:11:02,400
existing models and what have you and existing tools and make them their own.

154
00:11:02,400 --> 00:11:06,760
Just need to look at the Linux community and all the different flavors of Linux that exist

155
00:11:06,760 --> 00:11:08,960
in the world.

156
00:11:08,960 --> 00:11:14,440
Stable diffusion is a great example of this, where people have really fine tuned the models

157
00:11:14,440 --> 00:11:16,880
to get some really impressive results.

158
00:11:16,880 --> 00:11:19,600
There's a website that I use, a service called Super Machine.

159
00:11:19,600 --> 00:11:27,160
You can go to supermachine.art and they've taken a whole bunch of different use cases

160
00:11:27,160 --> 00:11:29,600
for stable diffusion and fine tuned it.

161
00:11:29,600 --> 00:11:31,960
So I'm just looking at a list of the options here.

162
00:11:31,960 --> 00:11:39,600
You've got super machine coloring book, which will create coloring book pictures, line drawings

163
00:11:39,600 --> 00:11:43,120
that look like they're great coloring books.

164
00:11:43,120 --> 00:11:51,080
You've got double exposure, low poly optimized versions, portrait plus, which is optimized

165
00:11:51,080 --> 00:11:54,080
for creating portraits of people.

166
00:11:54,080 --> 00:11:56,120
What's the other one?

167
00:11:56,120 --> 00:12:01,800
I was just looking at, oh yeah, wool ties, which is a model where all of the pictures

168
00:12:01,800 --> 00:12:08,520
look like whatever you ask for, we'll create it as if it's like knitted models.

169
00:12:08,520 --> 00:12:11,200
Everything's made of wool.

170
00:12:11,200 --> 00:12:17,240
But to take that to the extreme, you talk about the guardrails coming off.

171
00:12:17,240 --> 00:12:23,040
Most of these models don't allow you to create nudes or any kind of erotic imagery and things

172
00:12:23,040 --> 00:12:24,040
like that.

173
00:12:24,040 --> 00:12:30,740
But there is actually a website, nudes.ai, and I would heavily suggest that this is an

174
00:12:30,740 --> 00:12:33,080
not safe for work website.

175
00:12:33,080 --> 00:12:40,360
So don't fire up nudes.ai on your laptop going, oh yeah, Martin just suggested this website.

176
00:12:40,360 --> 00:12:48,960
To be clear, I'd never heard of it, but Martin knew a lot about it.

177
00:12:48,960 --> 00:12:53,240
He did, which is just an example that came up of what it looks like when you play with

178
00:12:53,240 --> 00:12:54,240
these models at the fringes.

179
00:12:54,240 --> 00:12:55,240
Definitely don't have them.

180
00:12:55,240 --> 00:13:00,240
I haven't absolutely rattled through and had to explain some awkward credit card purchases.

181
00:13:00,240 --> 00:13:03,480
All these credits, Martin, said my wife.

182
00:13:03,480 --> 00:13:06,000
No, I haven't actually in all seriousness.

183
00:13:06,000 --> 00:13:10,720
But to be fair, you don't really need to, you just go on the homepage and you get a

184
00:13:10,720 --> 00:13:15,680
very explicit view of exactly what nudes.ai is capable of.

185
00:13:15,680 --> 00:13:19,760
I believe that is a tuned model using stable diffusion.

186
00:13:19,760 --> 00:13:26,260
I may be wrong on that on the actual underlying model, but I'm pretty sure it's stable diffusion.

187
00:13:26,260 --> 00:13:33,640
So that also flags the challenges of launching these open source models, right?

188
00:13:33,640 --> 00:13:39,800
You basically are going to put that out there and there's going to be a percentage of people

189
00:13:39,800 --> 00:13:45,440
who lean into certain applications that could be way worse than what you just described.

190
00:13:45,440 --> 00:13:47,960
100% yeah.

191
00:13:47,960 --> 00:13:50,000
Right?

192
00:13:50,000 --> 00:13:51,000
That's a concern.

193
00:13:51,000 --> 00:13:58,480
And I think this is where the whole open or closed debate really rages for me is on those

194
00:13:58,480 --> 00:14:04,440
fringes because I think it's all well and good for, and this is just my personal opinion,

195
00:14:04,440 --> 00:14:10,560
for Stability AI to have this commitment to openness.

196
00:14:10,560 --> 00:14:13,880
But what are people going to do with it?

197
00:14:13,880 --> 00:14:20,280
And I just, I don't know if you can trust the 1% to not, to not, you know, do things

198
00:14:20,280 --> 00:14:23,240
that perhaps the 99% wish they hadn't.

199
00:14:23,240 --> 00:14:30,160
To be honest, you categorically cannot trust the 1% if people can do it, they will do it.

200
00:14:30,160 --> 00:14:31,160
They will push it.

201
00:14:31,160 --> 00:14:33,040
There will be someone that will push it to the limit.

202
00:14:33,040 --> 00:14:39,320
I guess that's what I'm trying to say, Martin, is that yeah, there'll always be someone.

203
00:14:39,320 --> 00:14:44,760
I guess the Stability AI is saying it's not up to us to say what people can and can't

204
00:14:44,760 --> 00:14:49,200
do, which I guess in free democracies is kind of a key component of that.

205
00:14:49,200 --> 00:14:55,440
But at the same time, yeah, probably outside the realms of this podcast for us to go into

206
00:14:55,440 --> 00:14:56,960
too much depth into that.

207
00:14:56,960 --> 00:15:03,920
But yeah, that Stability AI and their LLM and the balance of having a bit more flexibility

208
00:15:03,920 --> 00:15:07,120
with open source models, but some of the risks of open sourced models.

209
00:15:07,120 --> 00:15:11,720
If anyone is interested in that, just keep an eye out there, drip feeding the models

210
00:15:11,720 --> 00:15:14,840
as we speak, they're doing the smaller parameter model.

211
00:15:14,840 --> 00:15:19,160
So I think they've released an 8 billion parameter model at the moment and over the coming weeks

212
00:15:19,160 --> 00:15:25,640
and months, they'll be releasing the larger models over time.

213
00:15:25,640 --> 00:15:27,720
Good stuff.

214
00:15:27,720 --> 00:15:35,280
Let's move on to Google AI developments and this announcement of this word that I can't

215
00:15:35,280 --> 00:15:36,280
say, Martin.

216
00:15:36,280 --> 00:15:38,960
So why don't you tell us what the next story is?

217
00:15:38,960 --> 00:15:43,400
This is part of a wider piece from Google.

218
00:15:43,400 --> 00:15:44,920
They have two big announcements.

219
00:15:44,920 --> 00:15:47,840
Actually, only one of them is really an announcement.

220
00:15:47,840 --> 00:15:52,880
The other one is just kind of speculation, I guess, at this point.

221
00:15:52,880 --> 00:16:00,000
So the thing that you were talking about Project Magi, it's one of those two, I suspect someone's

222
00:16:00,000 --> 00:16:03,880
going to write in and tell us that we've got it wrong in both of those instances there.

223
00:16:03,880 --> 00:16:09,520
But Project Magi, according to the New York Times, Google has 160 engineers working on

224
00:16:09,520 --> 00:16:10,640
this project.

225
00:16:10,640 --> 00:16:16,360
And basically, it is the deeper integration of AI into search.

226
00:16:16,360 --> 00:16:23,640
So this is Google going, holy mackerel, we need to respond to chat GPT and chat GPT plugins

227
00:16:23,640 --> 00:16:25,560
and all of that kind of good stuff.

228
00:16:25,560 --> 00:16:31,960
And it's expected that there's going to be something more akin to a Google Bard, the

229
00:16:31,960 --> 00:16:39,640
chat interface coming into search, where we'll have less reliance of the kind of classic

230
00:16:39,640 --> 00:16:45,320
search engine results page with 10 organic listings and a bunch of other features like

231
00:16:45,320 --> 00:16:46,320
featured snippets.

232
00:16:46,320 --> 00:16:52,400
And that will be in there, but it's going to be much more chat oriented.

233
00:16:52,400 --> 00:16:55,200
We can ask follow up questions.

234
00:16:55,200 --> 00:16:59,020
But also there's some interesting capabilities that are being talked about, such as the ability

235
00:16:59,020 --> 00:17:01,480
to execute transactions.

236
00:17:01,480 --> 00:17:07,720
So you'll actually be able to make purchases through this new system.

237
00:17:07,720 --> 00:17:09,320
Now at the moment, it's all speculation.

238
00:17:09,320 --> 00:17:14,880
So there isn't a huge amount of confirmed detail other than the New York Times, which

239
00:17:14,880 --> 00:17:21,000
going to assume is fairly reliable, says that there is a resource of 160 people working

240
00:17:21,000 --> 00:17:22,480
on implementing this.

241
00:17:22,480 --> 00:17:27,920
Now I said this is one of two developments regarding Google and AI.

242
00:17:27,920 --> 00:17:31,000
And the other one was an announcement yesterday.

243
00:17:31,000 --> 00:17:41,960
So Google and DeepMind, so the Google Brain team, which was the AI research lab of Google

244
00:17:41,960 --> 00:17:52,040
and DeepMind, which is the UK based AI research lab owned by Google's parent company, Alphabet,

245
00:17:52,040 --> 00:17:54,520
they are becoming one.

246
00:17:54,520 --> 00:17:58,320
And they are now going to be Google DeepMind.

247
00:17:58,320 --> 00:18:05,080
And the idea is that they will be developing more advanced and versatile AIs.

248
00:18:05,080 --> 00:18:10,480
One of the areas that DeepMind has been focused on over the years is more scientific applications.

249
00:18:10,480 --> 00:18:14,720
So if you've been following the kind of protein folding, which I'm guessing you've seen a

250
00:18:14,720 --> 00:18:17,480
couple of articles about in your time, Paul.

251
00:18:17,480 --> 00:18:20,080
Love a bit of protein folding, mate.

252
00:18:20,080 --> 00:18:23,160
DeepMind has done a huge amount of work in that area.

253
00:18:23,160 --> 00:18:28,640
And they're very much interested in the science capabilities.

254
00:18:28,640 --> 00:18:37,600
If you watched AlphaGo, a Netflix documentary about the AI model that basically dominated

255
00:18:37,600 --> 00:18:44,080
the game of Go, that was all the DeepMind project as well.

256
00:18:44,080 --> 00:18:48,680
The idea being that bringing these two units together, Google Brain and DeepMind, you're

257
00:18:48,680 --> 00:18:52,080
just going to have economies of scale and you're going to have the best people working

258
00:18:52,080 --> 00:18:57,720
on the best kit, which is ultimately going to enable better collaboration and better

259
00:18:57,720 --> 00:19:02,700
speed of production and better outputs ultimately.

260
00:19:02,700 --> 00:19:08,760
In the announcement on the DeepMind blog, they said that there's going to be a new scientific

261
00:19:08,760 --> 00:19:13,980
board in charge of oversight as well.

262
00:19:13,980 --> 00:19:15,880
Which doesn't come as a huge surprise.

263
00:19:15,880 --> 00:19:18,640
DeepMind at one stage was...

264
00:19:18,640 --> 00:19:26,080
There's been concerns about DeepMind or within DeepMind about AI and how Google may or may

265
00:19:26,080 --> 00:19:27,240
not use it.

266
00:19:27,240 --> 00:19:33,800
There were reports of being concerned about Google's work with organisations like the

267
00:19:33,800 --> 00:19:38,880
military, kind of hardware, weapon-based companies and things like that.

268
00:19:38,880 --> 00:19:42,440
So I would imagine that this new scientific oversight board...

269
00:19:42,440 --> 00:19:45,320
And there'll be some compromises made all around, I'm sure.

270
00:19:45,320 --> 00:19:49,800
But ultimately, I suspect that this was a decision that was more or less forced upon

271
00:19:49,800 --> 00:19:50,800
DeepMind.

272
00:19:50,800 --> 00:19:54,200
Because I don't know that they came willingly.

273
00:19:54,200 --> 00:20:00,200
Yeah, I mean, a couple of things.

274
00:20:00,200 --> 00:20:03,200
I think we'll get to in a moment what this means for marketers.

275
00:20:03,200 --> 00:20:09,200
But there's enough there in both those stories, like Google is deeply unsettled.

276
00:20:09,200 --> 00:20:14,040
They are smashing together teams that may not want to be smashed together by the sounds

277
00:20:14,040 --> 00:20:16,040
of things.

278
00:20:16,040 --> 00:20:21,280
There was a news item, I can't remember if it was this week or last week, where there

279
00:20:21,280 --> 00:20:28,160
was a rumour that Samsung would be pulling their deal with Google, which they use on

280
00:20:28,160 --> 00:20:31,600
all their phones and replacing it with Bing.

281
00:20:31,600 --> 00:20:34,560
And it was like a $3 billion deal or something crazy.

282
00:20:34,560 --> 00:20:37,060
That again was prompting...

283
00:20:37,060 --> 00:20:38,800
Did you see what that did to the market cap?

284
00:20:38,800 --> 00:20:39,800
What happened?

285
00:20:39,800 --> 00:20:40,800
Was there a drop?

286
00:20:40,800 --> 00:20:43,640
There's a $60 billion drop.

287
00:20:43,640 --> 00:20:50,400
So Google is struggling a little bit, would be the inference from some of this stuff.

288
00:20:50,400 --> 00:21:00,200
In terms of what it means for marketers per se, looking at Magi, I think getting closer,

289
00:21:00,200 --> 00:21:03,760
maybe I'm getting closer.

290
00:21:03,760 --> 00:21:06,240
What would this mean for marketers?

291
00:21:06,240 --> 00:21:11,760
I have read online that it's going to bring additional features to search, but the Google

292
00:21:11,760 --> 00:21:13,360
ads are still going to be a thing.

293
00:21:13,360 --> 00:21:17,280
So advertising is still going to be part of our search results page.

294
00:21:17,280 --> 00:21:21,400
Obviously you mentioned the direct buy.

295
00:21:21,400 --> 00:21:26,040
There's I assume going to be some sort of mechanism by which the person selling the

296
00:21:26,040 --> 00:21:31,240
thing obviously still gets the cash, but Google will get a share of that by enabling that.

297
00:21:31,240 --> 00:21:37,880
And therefore as an e-commerce marketer or some other marketer that's maybe delivering

298
00:21:37,880 --> 00:21:41,680
information products or other things where that might be a good way for you to generate

299
00:21:41,680 --> 00:21:42,680
revenue.

300
00:21:42,680 --> 00:21:45,400
But again, is that now an extra cost for you in terms of, is that going to be more than

301
00:21:45,400 --> 00:21:47,200
you stand at Google ad?

302
00:21:47,200 --> 00:21:50,160
Because now we're not just talking about a click, right?

303
00:21:50,160 --> 00:21:54,920
We're talking about a conversion where someone spends money and therefore would Google expect

304
00:21:54,920 --> 00:21:59,040
to take a bigger share of that than they have been with clicks?

305
00:21:59,040 --> 00:22:02,080
How fast is this going to happen would be a key question as well.

306
00:22:02,080 --> 00:22:06,480
And I've, I saw an article in search engine journal that suggests that Google's going

307
00:22:06,480 --> 00:22:11,920
to introduce this to 1 million users initially, and that that will grow to 30 million users

308
00:22:11,920 --> 00:22:18,880
by the end of the year and it will be accessible only in the US to begin with.

309
00:22:18,880 --> 00:22:25,080
So certainly it's not, it's going to be a phased rollout and it's not going to totally

310
00:22:25,080 --> 00:22:28,120
disrupt how we all use search to begin with.

311
00:22:28,120 --> 00:22:32,720
But yes, if you're using, if you're doing search engine optimization, if you're using Google

312
00:22:32,720 --> 00:22:37,680
ads, this is going to have an impact on that ecosystem one assumes, right?

313
00:22:37,680 --> 00:22:43,840
Yeah, the search engine optimization thing is really interesting.

314
00:22:43,840 --> 00:22:51,600
When all of the transaction can happen within the search engine, that changes the game entirely.

315
00:22:51,600 --> 00:22:54,720
That's going to make Amazon sit up and listen as well.

316
00:22:54,720 --> 00:22:59,020
That's, they are, they are going to be watching closely.

317
00:22:59,020 --> 00:23:03,240
That's going to be taking away traffic to them for sure.

318
00:23:03,240 --> 00:23:12,920
So this is, this potentially is very disruptive, not just for vendors, but for other marketplaces

319
00:23:12,920 --> 00:23:13,920
as well.

320
00:23:13,920 --> 00:23:17,240
You know, what does this mean for the likes of Shopify?

321
00:23:17,240 --> 00:23:21,720
You know, is it going to be taking away traffic for them?

322
00:23:21,720 --> 00:23:26,500
Will they actually just integrate deeply into this and make it so easy for their customers

323
00:23:26,500 --> 00:23:32,760
to sell via whatever Google's new search transaction interface looks like that people actually

324
00:23:32,760 --> 00:23:36,680
stick with Shopify and it all works beautifully?

325
00:23:36,680 --> 00:23:38,120
I don't know.

326
00:23:38,120 --> 00:23:44,120
Suffice to say that if you're in e-commerce, pay attention to these upcoming changes.

327
00:23:44,120 --> 00:23:47,680
The AdWords thing's very interesting as well, isn't it?

328
00:23:47,680 --> 00:23:50,600
They, this is a, this is Google's cash cow, right?

329
00:23:50,600 --> 00:23:52,640
Google makes, what is it?

330
00:23:52,640 --> 00:23:59,920
Something like 200 odd billion a year in AdWords, through AdWords.

331
00:23:59,920 --> 00:24:04,320
So they can't take away that, they can't cannibalize that.

332
00:24:04,320 --> 00:24:09,840
So it has to feature and how they integrate that into the product.

333
00:24:09,840 --> 00:24:19,160
As we discussed before, this, the AI revolution that we're seeing at the moment is really

334
00:24:19,160 --> 00:24:25,620
a question about UI and user interfaces and just looking at the way that people are engaging

335
00:24:25,620 --> 00:24:33,120
with these new, these new pieces of tech, we've just talked about that search, but actually,

336
00:24:33,120 --> 00:24:36,360
and we'll come onto it in a minute, with Snapchat.

337
00:24:36,360 --> 00:24:40,400
These avatars are changing the way people interface with computers.

338
00:24:40,400 --> 00:24:44,120
So the whole thing is kind of up for grabs at the minute and anything that we knew as,

339
00:24:44,120 --> 00:24:47,720
well, anything that we maybe took for granted as being the way things were done over the

340
00:24:47,720 --> 00:24:51,040
past 20 years, I think is up for grabs.

341
00:24:51,040 --> 00:24:53,560
Yeah, I completely agree.

342
00:24:53,560 --> 00:24:58,240
I was reading in an article, I can't remember which one it was, but in essence, I think

343
00:24:58,240 --> 00:25:03,600
the quote was, Microsoft smells blood in the water, right?

344
00:25:03,600 --> 00:25:10,560
Bringing and infusing chat GPT, GPT-4 into as many of its technologies as possible because

345
00:25:10,560 --> 00:25:17,240
it lost out decades ago to Google in the battle for search and now, and has had to wait all

346
00:25:17,240 --> 00:25:22,120
those years for an opportunity to try and turn the tide and rebalance things.

347
00:25:22,120 --> 00:25:26,920
And I think this is all disruptive enough, as you're saying, Martin, to enable that.

348
00:25:26,920 --> 00:25:29,880
So search and mobile, they were behind on mobile as well.

349
00:25:29,880 --> 00:25:30,880
Yeah.

350
00:25:30,880 --> 00:25:32,800
So this is going to be interesting.

351
00:25:32,800 --> 00:25:36,880
I think your comment on Amazon is also an interesting one, right?

352
00:25:36,880 --> 00:25:43,120
Because we talked about the Amazon LLM release last week, which was not related to search

353
00:25:43,120 --> 00:25:47,960
for obvious reasons, because Amazon doesn't, I guess Amazon does have a search tool, but

354
00:25:47,960 --> 00:25:52,240
it's within its product driven landscape of what you're on Amazon looking for products.

355
00:25:52,240 --> 00:25:55,200
But when I searched for products on Google, Amazon is one of the key results that often

356
00:25:55,200 --> 00:25:58,040
comes up for obvious reasons.

357
00:25:58,040 --> 00:25:59,280
So what happens now?

358
00:25:59,280 --> 00:26:03,440
Does Amazon work really hard to try and make sure that the purchase that you make directly

359
00:26:03,440 --> 00:26:07,440
through the search engine still happens through Amazon?

360
00:26:07,440 --> 00:26:09,040
How's that all going to play out?

361
00:26:09,040 --> 00:26:13,880
It's, yeah, I think it's a wait and see because very, very hard to predict, especially until

362
00:26:13,880 --> 00:26:18,120
we see it out in the wild and how well it works and all that good stuff.

363
00:26:18,120 --> 00:26:20,120
But cool, cool, cool.

364
00:26:20,120 --> 00:26:21,120
Right.

365
00:26:21,120 --> 00:26:23,680
Let's move on to our next story then.

366
00:26:23,680 --> 00:26:28,360
This is Atlassian, which I always feel like I pronounce incorrectly.

367
00:26:28,360 --> 00:26:31,560
So maybe we're going to get some messages about that and people can educate me on how

368
00:26:31,560 --> 00:26:33,920
to pronounce that word.

369
00:26:33,920 --> 00:26:39,360
Makers of Confluence and Jira baking in AI into its tools.

370
00:26:39,360 --> 00:26:46,680
So this one, I think, at least as I understand it, I've never used Jira or Confluence, but

371
00:26:46,680 --> 00:26:55,440
I think these tend to be very popular with developers and people who use agile project

372
00:26:55,440 --> 00:26:56,640
management principles.

373
00:26:56,640 --> 00:27:02,480
But I have seen them more and more adopted by marketers, even in the agency world for

374
00:27:02,480 --> 00:27:09,440
those agencies operating in an agile way with point-based pricing systems and sprints and

375
00:27:09,440 --> 00:27:13,520
therefore adopting a more software development type mechanism.

376
00:27:13,520 --> 00:27:19,240
So for those of you marketers out there that are using these products to drive your project

377
00:27:19,240 --> 00:27:23,200
management, you'll get into AI augmentation.

378
00:27:23,200 --> 00:27:27,960
To give you some examples of what it can do, it can quickly summarize decisions and actions

379
00:27:27,960 --> 00:27:33,220
from meeting minutes, it can help the customer service aspects and the knowledge base aspects

380
00:27:33,220 --> 00:27:39,240
can help resolve help requests instantly based on understanding the knowledge base articles

381
00:27:39,240 --> 00:27:43,860
and the system can ask follow-up questions and take any necessary actions.

382
00:27:43,860 --> 00:27:49,100
It can help craft thoughtful responses to put customers at ease during stressful situations.

383
00:27:49,100 --> 00:27:56,560
It can generate insights using data from multiple sources in the Atlassian Analytics Cloud Hub

384
00:27:56,560 --> 00:28:01,600
to analyze and visualize data across all of Atlassian's products and even connected third-party

385
00:28:01,600 --> 00:28:06,520
tools without you needing to know how to write SQL or any other developer magic in order

386
00:28:06,520 --> 00:28:08,280
to be able to do that.

387
00:28:08,280 --> 00:28:10,680
So what does this mean for marketers?

388
00:28:10,680 --> 00:28:14,600
Well obviously if you're a user of those tools, you're about to get this augmentation, but

389
00:28:14,600 --> 00:28:21,960
more broadly when is AI coming to your favorite customer service health desk tool, your favorite

390
00:28:21,960 --> 00:28:25,300
project management tool?

391
00:28:25,300 --> 00:28:29,800
Maybe it's time to start having a look at those public roadmaps for the tools that you

392
00:28:29,800 --> 00:28:35,200
use and get a feel for when you get an AI augmentation because if this is going to be

393
00:28:35,200 --> 00:28:43,240
able to help save you 10, 20, 30, 50% time on certain tasks, this could be enough to

394
00:28:43,240 --> 00:28:45,000
get you to switch.

395
00:28:45,000 --> 00:28:49,160
Now might be the time to evaluate the tools you're using and how fast they're going to

396
00:28:49,160 --> 00:28:54,480
bring AI into the mix and whether or not actually want to move to a tool that is already doing

397
00:28:54,480 --> 00:28:56,800
it or looks like it will do it faster.

398
00:28:56,800 --> 00:29:03,360
So that's why I thought this story was interesting Martin, not just for the users of Atlassian

399
00:29:03,360 --> 00:29:08,720
products but also for all the rest of the products are now in catch-up mode.

400
00:29:08,720 --> 00:29:11,320
What are your thoughts?

401
00:29:11,320 --> 00:29:17,160
Video looks very slick, looks very cool, it's very similar to Copilot isn't it, you know,

402
00:29:17,160 --> 00:29:21,480
works across that product suite, it's using your company's data.

403
00:29:21,480 --> 00:29:31,120
Yeah, very, very smart and the response from similar organisations is going to be key here.

404
00:29:31,120 --> 00:29:34,760
If you're using this for your customer service ticketing or if you're not using this for

405
00:29:34,760 --> 00:29:38,880
your customer service ticketing, this could be enough to get you to switch as you just

406
00:29:38,880 --> 00:29:44,160
said it's going to reduce the time spent dealing with incoming inquiries, you'll get quicker

407
00:29:44,160 --> 00:29:49,760
first time solves for customers.

408
00:29:49,760 --> 00:29:54,880
I think we are seeing this in the other big players as well though, we're both connected

409
00:29:54,880 --> 00:29:59,320
with the HubSpot ecosystem and I think it's difficult to go on LinkedIn at the moment

410
00:29:59,320 --> 00:30:07,000
without seeing Darmesh Shah talking about ChatSpot or what they are doing in the AI

411
00:30:07,000 --> 00:30:12,420
space and I think we're going to see very similar capabilities on a customer service

412
00:30:12,420 --> 00:30:16,760
front with the support hub in the near future.

413
00:30:16,760 --> 00:30:22,440
There were conversations about this on LinkedIn this week where senior product managers at

414
00:30:22,440 --> 00:30:26,920
HubSpot were basically saying, hey guys, let's give us some feedback about what you want

415
00:30:26,920 --> 00:30:28,440
to see in the tech stack.

416
00:30:28,440 --> 00:30:31,960
So I think everybody is aware that this is the way it is going and again this goes back

417
00:30:31,960 --> 00:30:37,240
to that custom embedding piece, the companies that can help you embed your company data

418
00:30:37,240 --> 00:30:42,200
into a model and make it accessible and make it readable are going to be the ones that

419
00:30:42,200 --> 00:30:47,360
add huge amounts of value on top of what is already very interesting technology in the

420
00:30:47,360 --> 00:30:51,480
foundational model, LLM space.

421
00:30:51,480 --> 00:30:56,520
So that's where companies have to be looking, go to your question about when is AI coming

422
00:30:56,520 --> 00:30:59,440
to your favourite PM tool.

423
00:30:59,440 --> 00:31:04,400
I've been looking at the project management tool landscape for answering this question

424
00:31:04,400 --> 00:31:11,960
recently and most of them have the capability to do some basic generation and text summarisation

425
00:31:11,960 --> 00:31:20,280
because they've integrated GPT 3.5 or GPT 4 or whatever model it is, but it's quite

426
00:31:20,280 --> 00:31:27,880
basic it's mostly on summarise these meeting notes, generate me a project plan and give

427
00:31:27,880 --> 00:31:31,920
me an agenda for this upcoming sprint meeting or what have you.

428
00:31:31,920 --> 00:31:36,480
The real power comes from taking your data and making that functional and usable and

429
00:31:36,480 --> 00:31:39,160
that's where you have to be looking.

430
00:31:39,160 --> 00:31:46,160
It's beyond just the kind of surface level generative AI that project management software

431
00:31:46,160 --> 00:31:47,240
has to be thinking about.

432
00:31:47,240 --> 00:31:56,320
Now it's about how do we take real world data, your emails, your meeting notes, your sprint

433
00:31:56,320 --> 00:32:02,360
Gantt charts, your code base, your knowledge base, everything and turn that into the model

434
00:32:02,360 --> 00:32:07,580
so every member of your team can access it and your customers can get insights from knowledge

435
00:32:07,580 --> 00:32:09,640
base and touch my button as well.

436
00:32:09,640 --> 00:32:16,320
Yeah I agree and I think as a marketer thinking about what data we already have, where is

437
00:32:16,320 --> 00:32:22,480
it stored, how is it labelled if appropriate or structured, although one hopes these tools

438
00:32:22,480 --> 00:32:27,320
improve their ability to pass unstructured data and figure out what it's about and how

439
00:32:27,320 --> 00:32:30,680
to use it and to answer different questions will increase.

440
00:32:30,680 --> 00:32:38,720
That goes to the point about the companies that make that easy for you, using vector

441
00:32:38,720 --> 00:32:46,060
embeddings, fine tuning LLMs, crafting an open source model to bend it to your will,

442
00:32:46,060 --> 00:32:53,320
this is hard stuff, you have to be techie to do that and most marketing teams aren't

443
00:32:53,320 --> 00:32:57,040
thinking about that.

444
00:32:57,040 --> 00:33:01,160
The tools that just basically say don't worry about that, we've got that, we've done that

445
00:33:01,160 --> 00:33:08,240
for you, we've turned your data into vector embeddings and fed that into a database using

446
00:33:08,240 --> 00:33:13,840
the pine cone API and integrated it with our product, you don't have to worry about any

447
00:33:13,840 --> 00:33:19,440
of that, all of that sounds very techie, very jargony, all you need to know is your data

448
00:33:19,440 --> 00:33:23,600
can now be searched, indexed and used as part of this AI.

449
00:33:23,600 --> 00:33:26,600
They're the companies that are going to win out.

450
00:33:26,600 --> 00:33:29,760
I completely agree.

451
00:33:29,760 --> 00:33:37,400
It also makes me think of, have you read the book Principles by Ray Dalio of Bridgewater

452
00:33:37,400 --> 00:33:38,400
Associates?

453
00:33:38,400 --> 00:33:40,840
I can't say I haven't.

454
00:33:40,840 --> 00:33:47,320
Yeah, it's a fascinating book and Ray Dalio is a fascinating person, at least to me.

455
00:33:47,320 --> 00:33:55,760
As part of that, he talks about how Bridgewater Associates as a hedge fund recorded every

456
00:33:55,760 --> 00:33:59,920
meeting.

457
00:33:59,920 --> 00:34:07,520
I'm talking for 20 years of data, not like we can now with the authors of the world or

458
00:34:07,520 --> 00:34:10,840
maybe even a lot of digital meetings like what we're having now.

459
00:34:10,840 --> 00:34:19,480
It really struck me that having that repository of information, if we're assuming that they're

460
00:34:19,480 --> 00:34:25,360
recording practically everything and they're having key conversations where they're discussing

461
00:34:25,360 --> 00:34:30,140
and exploring issues that the business is facing, they're looking at the key information

462
00:34:30,140 --> 00:34:36,880
points to infer into that and then they're making decisions.

463
00:34:36,880 --> 00:34:45,480
In essence, you could then train an AI using that data to make decisions like the management

464
00:34:45,480 --> 00:34:51,160
team of Bridgewater Associates would make decisions based on all of those recordings

465
00:34:51,160 --> 00:34:52,160
of how Bridgewater-

466
00:34:52,160 --> 00:34:54,520
In 20 years of historic training data.

467
00:34:54,520 --> 00:34:55,520
Right?

468
00:34:55,520 --> 00:34:59,520
A, that's a bit scary, but B, that's interesting.

469
00:34:59,520 --> 00:35:04,560
Certainly in my own mind, I'm really thinking how much more proactive do I want to be in

470
00:35:04,560 --> 00:35:10,240
recording the meetings and other mechanisms of communication that we have at Biostrata

471
00:35:10,240 --> 00:35:13,640
because maybe this is a piece of value in those meetings we want to be able to unleash

472
00:35:13,640 --> 00:35:15,040
at some point.

473
00:35:15,040 --> 00:35:23,760
Some of that you're going to get from being able to internalize and understand the sheer

474
00:35:23,760 --> 00:35:27,840
volume of email communication that goes on in organizations.

475
00:35:27,840 --> 00:35:31,760
I guess then that raises a question about how will the AI be able to tell what was a

476
00:35:31,760 --> 00:35:36,600
good decision, what was a bad decision and what was a useful conversation that ended

477
00:35:36,600 --> 00:35:40,880
up having an impact and what ended up being like a useless sideline.

478
00:35:40,880 --> 00:35:47,880
But all of those things I think end up being quite fascinating in terms of what data sources

479
00:35:47,880 --> 00:35:48,880
do you have?

480
00:35:48,880 --> 00:35:52,680
You know, Bridgewater Associates can really do an amazing thing for that.

481
00:35:52,680 --> 00:35:58,600
When I was listening to a podcast, sorry, I'm into an absolute rant now, today, where

482
00:35:58,600 --> 00:36:04,160
apologies, because my memory is so bad for names, I'm not going to attribute this properly,

483
00:36:04,160 --> 00:36:07,640
but we'll put it in the show notes.

484
00:36:07,640 --> 00:36:14,160
It was a podcast where they were talking about interesting ways to customize GPT-3 and train

485
00:36:14,160 --> 00:36:19,840
it and what someone had done is they'd fed it a load of their electronic journal notes,

486
00:36:19,840 --> 00:36:26,360
journaling for self-reflection and mindfulness and just, you know, as part of their own self-maintaining

487
00:36:26,360 --> 00:36:28,920
their own mental health.

488
00:36:28,920 --> 00:36:35,120
And this person fed it into GPT-3 and then it asked what their Myers-Briggs profile was

489
00:36:35,120 --> 00:36:40,080
based on that information and it got it right.

490
00:36:40,080 --> 00:36:41,080
Wow.

491
00:36:41,080 --> 00:36:45,320
Therefore, what other information could you glean?

492
00:36:45,320 --> 00:36:51,000
Now we're drifting into maybe like slightly scary George Orwell territory, but what other

493
00:36:51,000 --> 00:36:54,800
information could you glean about people if you had access to all their emails or their

494
00:36:54,800 --> 00:36:55,800
meetings?

495
00:36:55,800 --> 00:37:01,200
Well, on that note, there's a tool called Crystal Nose and that's Crystal Nose as in

496
00:37:01,200 --> 00:37:08,600
knowledge, not nose as in olfactory on the front of your face.

497
00:37:08,600 --> 00:37:17,080
CrystalNose.com, I've had a subscription to it, 2016, so this is not new tech by any stretch.

498
00:37:17,080 --> 00:37:22,120
It allows you to go on a LinkedIn profile and it gives you somebody's disc profile and

499
00:37:22,120 --> 00:37:24,580
it gives you sales coaching insights.

500
00:37:24,580 --> 00:37:29,440
It gives you like how you should write emails to them.

501
00:37:29,440 --> 00:37:30,440
It's incredible.

502
00:37:30,440 --> 00:37:38,120
I've had some really scary, like scarily accurate insights provided through this tool.

503
00:37:38,120 --> 00:37:43,360
This one that I came across on an inbound event a few years ago and actually it was

504
00:37:43,360 --> 00:37:48,080
Paul Raitzer from the Marketing Institute who mentioned it and I subscribed pretty much

505
00:37:48,080 --> 00:37:51,400
there and then and I've had an active subscription for years now.

506
00:37:51,400 --> 00:37:52,400
Wow.

507
00:37:52,400 --> 00:37:57,560
Well, they're going to have potentially even more data and more power to do that and it's

508
00:37:57,560 --> 00:38:02,320
something we said maybe in episode one or episode two when we got started on the podcast

509
00:38:02,320 --> 00:38:12,640
is with great power comes great responsibility and what is an ethical, practical way to leverage

510
00:38:12,640 --> 00:38:18,360
these tools for the benefit of everyone to improve effectiveness and efficiency but without

511
00:38:18,360 --> 00:38:26,920
drifting into icky areas of unfair and unethical profiling, for example.

512
00:38:26,920 --> 00:38:31,720
I didn't even mean for this to go down this avenue when I mentioned Bridgewater Associates

513
00:38:31,720 --> 00:38:37,200
and how cool it would be to have all of that meeting data to create a management team AI

514
00:38:37,200 --> 00:38:41,840
but I guess once you explore it deep enough you can really start to think about all the

515
00:38:41,840 --> 00:38:47,480
other different ways that you could use that data but yeah there's a bit of a digression

516
00:38:47,480 --> 00:38:51,120
folks hopefully it was still useful to hear Martin and I banter back and forth on that

517
00:38:51,120 --> 00:38:56,280
but probably in the interest of time I'm sure that your dog walk or gym visit or whatever

518
00:38:56,280 --> 00:38:58,440
you're doing is going to be over soon.

519
00:38:58,440 --> 00:39:01,280
We should probably move on to the next story Martin.

520
00:39:01,280 --> 00:39:07,280
Adobe teases Firefly's AI driven tech will be coming to video editing.

521
00:39:07,280 --> 00:39:16,080
They've published a video it's about 90 seconds long showcasing Firefly which is the text

522
00:39:16,080 --> 00:39:21,440
to image tool that was recently announced I think it feels like that was announced what

523
00:39:21,440 --> 00:39:24,280
three or four weeks ago.

524
00:39:24,280 --> 00:39:30,260
They've now basically said look this they've showcased that this can be used for video editing

525
00:39:30,260 --> 00:39:34,100
and it is so cool.

526
00:39:34,100 --> 00:39:39,600
You will upload your video and you can say things like change the videos color grade

527
00:39:39,600 --> 00:39:48,120
or change the background or you can get it to transcribe the text and then it will find

528
00:39:48,120 --> 00:39:53,800
b-roll footage to match the text of what the person's talking about so there's a clip where

529
00:39:53,800 --> 00:39:59,760
there's a it's like a fake advert for some trainers some running shoes and this woman's

530
00:39:59,760 --> 00:40:04,600
sat down talking about her trainers and as she's talking about them and what she does

531
00:40:04,600 --> 00:40:12,040
with them it it transcribes her tech or her commentary and then finds footage of a close

532
00:40:12,040 --> 00:40:16,960
up of the shoe of the shoe kind of running through the woods of basically doing what

533
00:40:16,960 --> 00:40:19,920
she's what she's talking about.

534
00:40:19,920 --> 00:40:31,000
It's very very cool also for sound editing so you can get it to to find relevant sound

535
00:40:31,000 --> 00:40:39,520
effects and the clip that it shows in the video is the sea the tide coming in you know

536
00:40:39,520 --> 00:40:46,240
coming up against the shoreline and it kind of you say find sound effects it comes up

537
00:40:46,240 --> 00:40:54,440
with waves or seaside kind of noises you select that it drops into the video.

538
00:40:54,440 --> 00:41:03,040
If you're doing video production this is going to speed up your your workflow no end.

539
00:41:03,040 --> 00:41:11,340
It looks super cool converts scripts into detailed storyboards pre-visuals b-roll recommendations

540
00:41:11,340 --> 00:41:17,200
like I say automated color grading and a bunch of other tools that I'm sure someone like

541
00:41:17,200 --> 00:41:23,760
me who doesn't know anything about video editing would would use you know it is yeah and this

542
00:41:23,760 --> 00:41:28,120
is an exciting time and the great thing about the Adobe Firefly is that it's all trained

543
00:41:28,120 --> 00:41:34,120
on data which isn't copyright so if you're concerned about that I don't know a lot of

544
00:41:34,120 --> 00:41:44,120
companies are tentatively or not using some generative AI models because of that the copyright

545
00:41:44,120 --> 00:41:51,080
issues for example Wired magazine isn't using generative AI text to image until the copyright

546
00:41:51,080 --> 00:41:53,080
issue is sorted out.

547
00:41:53,080 --> 00:42:00,440
Yeah Firefly doesn't have that issue because it's all trained on data that Adobe can legitimately

548
00:42:00,440 --> 00:42:01,440
use.

549
00:42:01,440 --> 00:42:06,920
Right and and for for listeners that don't know this is driven by the fact that a lot

550
00:42:06,920 --> 00:42:15,400
of the training of other models has been on copyrighted images and we're now seeing lawsuits

551
00:42:15,400 --> 00:42:22,240
and other legal challenges to open AI for DORLII 2 and stable diffusion and all these

552
00:42:22,240 --> 00:42:28,240
other models basically saying you didn't pay us to access these large imagery image repositories

553
00:42:28,240 --> 00:42:34,600
to do this training especially now your model can use those images almost as inspiration

554
00:42:34,600 --> 00:42:39,720
for creating new images but in this case Adobe trained it on a bunch of images that were

555
00:42:39,720 --> 00:42:44,560
already it already owned the copyright to do so.

556
00:42:44,560 --> 00:42:49,240
Really interesting I I get really excited about these tools like I just think about

557
00:42:49,240 --> 00:42:55,560
playing with them I think if you are a professional videographer and you do this type of work

558
00:42:55,560 --> 00:43:00,720
it was back to the old adage that's I say old adage it's probably been only in common

559
00:43:00,720 --> 00:43:04,200
use for like two months but I've just heard it so many times now I never really want to

560
00:43:04,200 --> 00:43:07,400
hear it again so sorry I'm about to repeat it for all of you that have heard it a thousand

561
00:43:07,400 --> 00:43:13,680
times yourselves but that videographer AI won't replace videographers but videographers

562
00:43:13,680 --> 00:43:18,520
that use AI will replace those that don't and I think this is another case of that.

563
00:43:18,520 --> 00:43:21,400
One thing that was related to this I had a really interesting conversation with a good

564
00:43:21,400 --> 00:43:27,300
friend of mine Giovanni who I know is a listener to the podcast and is a incredibly smart person

565
00:43:27,300 --> 00:43:31,560
and is doing some really interesting things with it with AI himself.

566
00:43:31,560 --> 00:43:37,480
We were talking about Microsoft Copilot and how sexy and awesome that demo video is and

567
00:43:37,480 --> 00:43:43,040
look all this cool stuff you're going to do and now we've seen Firefly's amazing text

568
00:43:43,040 --> 00:43:50,120
to video edit cleverness and Giovanni said to me yeah but Paul don't you feel like there

569
00:43:50,120 --> 00:43:55,480
are things in Microsoft tech stack that have like not worked very well for like a decade

570
00:43:55,480 --> 00:44:01,600
right like SharePoint it's got better but it's still like buggy right Teams is now better

571
00:44:01,600 --> 00:44:08,040
but it had long buggy periods like these videos show us the absolute best case scenario of

572
00:44:08,040 --> 00:44:10,680
how these tools are going to work.

573
00:44:10,680 --> 00:44:17,680
How long will it take in the case of Copilot let's say Microsoft to embed them in a fully

574
00:44:17,680 --> 00:44:22,160
functioning format that doesn't break often enough to make you just want to smash it up

575
00:44:22,160 --> 00:44:25,720
and that's a question that we can't answer yet until we get to our hands on it and start

576
00:44:25,720 --> 00:44:32,160
playing with it but I suspect that they won't be as powerful and as seamless as the videos

577
00:44:32,160 --> 00:44:33,160
make you know.

578
00:44:33,160 --> 00:44:38,560
Absolutely it's a it's such a fair point and you know these are the the movie trailers

579
00:44:38,560 --> 00:44:42,000
of software aren't they and they show you the best bits and they're incredibly well

580
00:44:42,000 --> 00:44:49,280
edited and they're not going to show you the bit where it is recommending woodland noises

581
00:44:49,280 --> 00:44:55,920
for a sports arena event and you're just like what that makes no sense yeah.

582
00:44:55,920 --> 00:45:01,080
There's a lot of promise the reality remains to be seen but even so you know we know that

583
00:45:01,080 --> 00:45:05,640
certainly in my day-to-day workflow there's now so many little bits of AI that are giving

584
00:45:05,640 --> 00:45:11,680
me incremental gains that I'm wildly more efficient and yes it doesn't all work seamlessly

585
00:45:11,680 --> 00:45:16,000
all of the time at all I was using this morning I found a little bit frustrating I won't go

586
00:45:16,000 --> 00:45:23,720
into that today but most of the time I'm getting a little bit quicker and as I learn to use

587
00:45:23,720 --> 00:45:28,120
when the product you know when you when you start to work with the product a little bit

588
00:45:28,120 --> 00:45:32,960
you know its limits you know its capabilities you know when to deploy it so I think there

589
00:45:32,960 --> 00:45:37,760
will be an element of the trailer says it can do all of these things the reality is

590
00:45:37,760 --> 00:45:44,680
that when you use it it has some quirks but actually in the end it will make you a better

591
00:45:44,680 --> 00:45:49,120
video or more efficient videographer.

592
00:45:49,120 --> 00:45:54,640
Yeah yeah I think you're right cool right we've got quite three more bits we want to

593
00:45:54,640 --> 00:46:00,000
get through I think we rattle through these before we run out of time so let's talk Snapchat

594
00:46:00,000 --> 00:46:08,160
very briefly so Snapchat's AI chatbot has gone pretty viral this week I was reading

595
00:46:08,160 --> 00:46:14,300
one of the great newsletters that I subscribed to that highlighted that premium users have

596
00:46:14,300 --> 00:46:19,280
already been sending two million messages a day to their chatbots and that they're

597
00:46:19,280 --> 00:46:29,120
about to roll this out to more users basically every kid with an iPhone and that the uptake

598
00:46:29,120 --> 00:46:33,600
of it has been really massive and I think there's a number of things here that we've

599
00:46:33,600 --> 00:46:36,920
given that we've already touched on a number of controversial topics today we won't touch

600
00:46:36,920 --> 00:46:43,480
on the concept of letting children interact with these chatbots and how well aligned are

601
00:46:43,480 --> 00:46:47,320
they and how well controlled are they and what type of things are they going to be saying

602
00:46:47,320 --> 00:46:52,040
to kids that we prefer that they didn't but if you do have kids and your kids do use Snapchat

603
00:46:52,040 --> 00:46:59,080
I would probably go diving into this if I were you but just more broadly this friendly

604
00:46:59,080 --> 00:47:05,280
conversation with chatbots is an interesting one right there's other tools like this with

605
00:47:05,280 --> 00:47:13,640
like replica and character we've talked about before Martin that's the head honchos are

606
00:47:13,640 --> 00:47:19,820
some of the the royal family of the of the AI world are in charge of and is attracting

607
00:47:19,820 --> 00:47:25,280
a lot of interest and a lot of money where they're creating different characters that

608
00:47:25,280 --> 00:47:31,960
you can converse with and in characters case it's you can speak with like Donald Trump

609
00:47:31,960 --> 00:47:37,840
or Elon Musk or Socrates or you know whoever you can imagine and then they've trained them

610
00:47:37,840 --> 00:47:43,160
or customized each bot to kind of respond and act like that character but I think the

611
00:47:43,160 --> 00:47:48,440
thing that's proving interesting to me about this is especially for younger generations

612
00:47:48,440 --> 00:47:53,680
that are very tech savvy their willingness potentially to embrace this technology and

613
00:47:53,680 --> 00:47:59,600
have conversations with non-humans and enjoy them and embrace them and want to have them

614
00:47:59,600 --> 00:48:05,880
like they were having them with humans right there's been a lot of talk about therapy bots

615
00:48:05,880 --> 00:48:11,760
right and I was playing with a with a WhatsApp bot that in essence was an effectively like

616
00:48:11,760 --> 00:48:15,520
a mindfulness tool and would ask you what you're grateful for today and a lot of what

617
00:48:15,520 --> 00:48:22,000
these mindful journals do I'd always thought people won't use them right they'll know it's

618
00:48:22,000 --> 00:48:27,040
a robot they it won't be interesting and it will feel fake but so far behaviors perhaps

619
00:48:27,040 --> 00:48:32,560
suggesting otherwise and it maybe wouldn't take us that long to get used to it and accept

620
00:48:32,560 --> 00:48:37,680
it and it for you know for us all to have a personal psychologist or you know a shrink

621
00:48:37,680 --> 00:48:42,560
in a pot in our pocket who there to help us provide us with empathy and ask us great questions

622
00:48:42,560 --> 00:48:47,200
to help us process our own thinking perhaps that's why these companies are getting so

623
00:48:47,200 --> 00:48:52,120
much money because actually they're predicting that we will embrace this stuff it all comes

624
00:48:52,120 --> 00:48:57,840
back to user interface you know these these AI language models have been around for a

625
00:48:57,840 --> 00:49:02,760
number of years the fact that you could interact with chat GPT changed the game the underlying

626
00:49:02,760 --> 00:49:07,360
model I didn't change the way you interacted it with it had changed the game and when you

627
00:49:07,360 --> 00:49:15,640
plug that into characters and avatars that have faces and voices and they interact you

628
00:49:15,640 --> 00:49:21,720
know video games people love interacting with video games and you know NPCs and non-playable

629
00:49:21,720 --> 00:49:27,640
characters in video games are not particularly dynamic yet we still spend hours and hours

630
00:49:27,640 --> 00:49:35,000
and hours interacting in worlds populated by them when you can have a chat bot on your

631
00:49:35,000 --> 00:49:41,960
phone it gives you advice that listens to you that is engaged in the conversation yeah it's

632
00:49:41,960 --> 00:49:48,520
not hard to see why a younger generation that grows up with them is is going to to see that

633
00:49:48,520 --> 00:49:53,520
as the new normal yeah I mean I agree and even just as thinking about it now as we talk

634
00:49:53,520 --> 00:50:00,080
about it a conversational agent that gives you 100 percent of its time and attention

635
00:50:00,080 --> 00:50:04,240
how easy is that to get from other people in the real world these days right we're all

636
00:50:04,240 --> 00:50:08,720
in conversations where we've got a phone in our hand it's like oh sorry Martin I just

637
00:50:08,720 --> 00:50:13,680
got a whatsapp notification that's absolutely worth none of my time but I'm so addicted

638
00:50:13,680 --> 00:50:19,160
to this phone in my hand I will have to stop our conversation to look at it but now we've

639
00:50:19,160 --> 00:50:24,040
got a bot that won't do that right it'll pay attention to us so it's it's a fascinating

640
00:50:24,040 --> 00:50:28,640
one we mentioned the the idea of having it as like a therapy bot or something like that

641
00:50:28,640 --> 00:50:33,360
a key concept that you come across in in therapy if you're training to be a counselor or something

642
00:50:33,360 --> 00:50:39,400
like that is the therapeutic hour and it's you know 50 minutes long then that's not going

643
00:50:39,400 --> 00:50:43,240
to be the case anymore there's not going to be someone sat on a couch opposite you going

644
00:50:43,240 --> 00:50:48,720
well that's enough for one day I assume the problem solved now because we're at 48 minutes

645
00:50:48,720 --> 00:50:56,000
so yeah no you're absolutely right so apart from the fact I think this is a fascinating

646
00:50:56,000 --> 00:51:01,080
interrogation of social dynamics and how will people interact with technology as you very

647
00:51:01,080 --> 00:51:07,000
rightly said Martin now that they've broken down or almost come up with this new way of

648
00:51:07,000 --> 00:51:14,280
interacting with these models is as a marketer how could you capitalize on this what niche

649
00:51:14,280 --> 00:51:22,320
that bot can you create to support your customers perhaps based on data that you've already

650
00:51:22,320 --> 00:51:28,760
got or data that you could collect so that they had your brand in their pocket and they

651
00:51:28,760 --> 00:51:35,480
turn to your brand when solving certain work if you're in b2b certain work issues or having

652
00:51:35,480 --> 00:51:41,400
certain work questions or if you're in b2c how can you augment your relationship with

653
00:51:41,400 --> 00:51:46,840
your customers by being more than just a product to them but to you know make yourself a part

654
00:51:46,840 --> 00:51:54,080
of their life in other ways so I have no idea the scope of what those things could take

655
00:51:54,080 --> 00:51:59,040
the form of but I think especially if I was a bigger company I'd want a team thinking

656
00:51:59,040 --> 00:52:04,160
about this honestly yeah and depending on your brand I can certainly think of some big

657
00:52:04,160 --> 00:52:10,080
brands that might even want to integrate with nudes.ai for their avatar you had to bring

658
00:52:10,080 --> 00:52:15,300
back a nudes.ai I've been sat here thinking am I gonna have to cut the nudes.ai bit and

659
00:52:15,300 --> 00:52:22,120
now fundamentally whether I want to or not no it has to stay it nips on right moving

660
00:52:22,120 --> 00:52:28,080
swiftly on tell us about this economist article that you found because it was fascinating.

661
00:52:28,080 --> 00:52:34,520
This week's economist has a AI special edition there are three special reports within it

662
00:52:34,520 --> 00:52:38,440
one of which by the way is all about large language models how they work the underlying

663
00:52:38,440 --> 00:52:43,320
technology and it's a really really comprehensive explainer so if you're wondering how vector

664
00:52:43,320 --> 00:52:50,720
embeddings and LLMs work check it out give it a read really really worth your while but

665
00:52:50,720 --> 00:52:57,480
in this article Gary Marcus and Ankur Rieul I think that's how you say I've almost certainly

666
00:52:57,480 --> 00:53:03,080
bodged it anyway we know that pronunciation isn't our forte on this podcast they propose

667
00:53:03,080 --> 00:53:11,720
the creation of an international agency for AI the IAAI and the idea being that this is

668
00:53:11,720 --> 00:53:18,880
a global neutral not-for-profit organization that addresses the risks and challenges that

669
00:53:18,880 --> 00:53:25,480
AI poses and tries to establish some ethical guidelines and a set of standards for further

670
00:53:25,480 --> 00:53:31,000
AI development which off the back of the letter that we had recently calling for a hiatus

671
00:53:31,000 --> 00:53:38,720
on AI development it's certainly a hot topic you can see why this is being called for just

672
00:53:38,720 --> 00:53:43,640
on that I noticed that Elon Musk did sign that letter and then this week announced he's

673
00:53:43,640 --> 00:53:51,960
creating his own truth GPT an AI company so I saw that I saw a great meme that was like

674
00:53:51,960 --> 00:54:01,200
it was a play on Star Wars where Emperor Palpatine is also within the Senate as like potential

675
00:54:01,200 --> 00:54:07,120
goodie but was also obviously the Emperor as well where it was like let's the good version

676
00:54:07,120 --> 00:54:11,640
was like let's put a pause on AI and then the dark version was like let's create our

677
00:54:11,640 --> 00:54:17,400
own AI company and it was like hmm perhaps we were not far off in our supposition and

678
00:54:17,400 --> 00:54:19,320
a previous episode there Martin.

679
00:54:19,320 --> 00:54:28,320
Anyway enough about that man and his failed rocket.

680
00:54:28,320 --> 00:54:35,800
So yeah well back to the IAAI the idea being that this organization would also encourage

681
00:54:35,800 --> 00:54:40,640
international cooperation and knowledge sharing personally I don't think that's a massive

682
00:54:40,640 --> 00:54:46,400
issue the knowledge sharing at the moment just scratched the surface of the AI ML community

683
00:54:46,400 --> 00:54:52,280
and it's very clear to see that knowledge is being shared I think hugging face are doing

684
00:54:52,280 --> 00:54:59,200
a great job of enabling that but yeah they want to develop tech tools for auditing AI

685
00:54:59,200 --> 00:55:09,840
systems and curbing misinformation but critics are saying that the incentives are too great

686
00:55:09,840 --> 00:55:10,840
for that.

687
00:55:10,840 --> 00:55:16,560
So as we've discussed previously companies that win in this space are going to win for

688
00:55:16,560 --> 00:55:25,560
a long time so the competition is an AI dominance is just going to hinder the effectiveness

689
00:55:25,560 --> 00:55:34,520
of such an organization at a geopolitical level I think where does this sit right that

690
00:55:34,520 --> 00:55:42,080
America is going to want to have AI systems that American companies control China is going

691
00:55:42,080 --> 00:55:46,840
to want to have AI systems that they can control and use Russia I'm sure is going to be developing

692
00:55:46,840 --> 00:55:51,480
their own and the EU is going to have a lot to say as well.

693
00:55:51,480 --> 00:55:57,200
Who is this supranational organization what are they going to do to enforce it if China

694
00:55:57,200 --> 00:56:02,760
steps out of line and says you know what we've just created an AGL and it's going to take

695
00:56:02,760 --> 00:56:10,560
out all people with asthma and that's what we've created for whatever reason who's going

696
00:56:10,560 --> 00:56:16,320
to stop them but this AIAAI by the way I don't think Xi Jinping is planning that I don't

697
00:56:16,320 --> 00:56:25,120
think he's got any great portents to take out the asthmatics hope not as a weasel myself

698
00:56:25,120 --> 00:56:31,400
yeah so I just I think that the it's a good idea I think we do need more collaboration

699
00:56:31,400 --> 00:56:38,800
at an international level but the idea of having something that is enforceable on a

700
00:56:38,800 --> 00:56:43,240
geopolitical level is somewhat fanciful.

701
00:56:43,240 --> 00:56:53,520
I think we talked previously about this and when the letter came out suggesting the six

702
00:56:53,520 --> 00:56:57,440
month hiatus and just how hard that would be to please for all the reasons that you

703
00:56:57,440 --> 00:57:03,880
said this is a step in how to attempt to do that and but all the same limitations that

704
00:57:03,880 --> 00:57:08,640
you have just identified I think are still there I think the stakes are just far too

705
00:57:08,640 --> 00:57:16,200
high to just push the darker side of humanity to say yeah but if they don't agree you know

706
00:57:16,200 --> 00:57:20,680
if they are it's like game theory right it's a little bit like I was going to say if you

707
00:57:20,680 --> 00:57:28,360
game theory it's you absolutely can't trust that that person is doing the costs on you

708
00:57:28,360 --> 00:57:35,200
of not operating you have to take an A player strategy yeah so like it's I think that's

709
00:57:35,200 --> 00:57:43,520
why all of these things are grey but is it the prisoners dilemma yeah sorry I'm talking

710
00:57:43,520 --> 00:57:47,280
over you no no don't worry tell us about the prisoners dilemma you'll probably do a better

711
00:57:47,280 --> 00:57:48,280
job than me.

712
00:57:48,280 --> 00:57:58,000
Yeah well you can either collaborate or you can collude or basically snitch so two prisoners

713
00:57:58,000 --> 00:58:05,840
you can you're being quizzed if you snitch on the other person and they don't snitch

714
00:58:05,840 --> 00:58:15,520
you get released and they get ten years if you both say nothing you can be out in one

715
00:58:15,520 --> 00:58:24,480
year each if you both snitch on each other you both get 20 years or what have you so

716
00:58:24,480 --> 00:58:29,480
it's about what strategy do you take and there's different versions I mean it's all about the

717
00:58:29,480 --> 00:58:36,760
reward and the punishments and the incentives for that but ultimately what you should be

718
00:58:36,760 --> 00:58:43,780
doing is betray the group basically isn't it yeah betray the group in the classic prisoners

719
00:58:43,780 --> 00:58:49,200
dilemma the individual gets the greatest payoff if they betray the group rather than cooperate

720
00:58:49,200 --> 00:58:54,480
so in the version you just described you would be better off saying that the other person

721
00:58:54,480 --> 00:58:59,000
was a snitch and hoping they decided to uphold their side of the bargain and not call you

722
00:58:59,000 --> 00:59:04,480
a snitch you know or not say that you sorry not call you not say that you were in on it

723
00:59:04,480 --> 00:59:09,080
and then basically they go to prison and you don't and it's and I think this is the fundamental

724
00:59:09,080 --> 00:59:15,240
problem with all of this is I think this is a prisoner's dilemma situation where everyone

725
00:59:15,240 --> 00:59:21,640
knows that it's in their best interest to plow on with AI even if they're part of this

726
00:59:21,640 --> 00:59:26,640
international consortium and on the surface everybody's like yeah no these are the rules

727
00:59:26,640 --> 00:59:31,200
this is how we should play the game there's just too much incentive to go yeah but when

728
00:59:31,200 --> 00:59:35,960
we get back to the office this is how we're really going to play the game right but it's

729
00:59:35,960 --> 00:59:40,560
a zero shot game isn't it it's a zero shot game and if you take the B player strategy

730
00:59:40,560 --> 00:59:47,480
and you decide to to collude and the other person snitches like you're done you're six

731
00:59:47,480 --> 00:59:53,900
one we're going to talk about anthropic and Claude's model in a moment in a funding pitch

732
00:59:53,900 --> 01:00:03,240
deck anthropic said companies that have an advantage in this LLM space by 2025 2026 are

733
01:00:03,240 --> 01:00:10,640
going to be uncatchable right the pace of development at that point the leaders are

734
01:00:10,640 --> 01:00:17,880
going to be beset so if you take a six-month hiatus and everyone else goes you're out

735
01:00:17,880 --> 01:00:23,500
of the race yeah I mean they've got an incentive to say that in their pitch deck and yet intuitively

736
01:00:23,500 --> 01:00:29,520
it does feel true given the rate of change and the commercial societal ranging impacts

737
01:00:29,520 --> 01:00:33,600
of this technology right I think we should probably move on to the tool of the week now

738
01:00:33,600 --> 01:00:39,720
before we do I am just going to say hashtag save the weasers and I had to go on mute because

739
01:00:39,720 --> 01:00:45,780
I was laughing so much after your suggestion there of yourself as a weasel I'm a fellow

740
01:00:45,780 --> 01:00:52,920
weasel as well so I'm a big fan of save the weasels let's talk anthropics Claude then

741
01:00:52,920 --> 01:00:55,880
for tool of the week and then we'll let these lovely people go.

742
01:00:55,880 --> 01:01:02,560
For the week anthropics Claude a large language model with a chat based interface similar

743
01:01:02,560 --> 01:01:13,360
to chat GPT and Google's bard and other systems like that it is designed to be helpful not

744
01:01:13,360 --> 01:01:17,680
harmful and there's another thing that they say as well but basically it's designed to

745
01:01:17,680 --> 01:01:25,600
be good right it's supposed to be a helpful assistant which doesn't well like I say it

746
01:01:25,600 --> 01:01:31,780
doesn't do any harm like morally good you mean rather than deliver a high quality outcome

747
01:01:31,780 --> 01:01:37,000
yes exactly that and I kind of quizzed it when I was using it I was like what does that

748
01:01:37,000 --> 01:01:47,480
even really mean and upon digging into it it's basically a constitutional AI so where

749
01:01:47,480 --> 01:01:54,600
other models such as chat GPT are trained on reinforcement learning through human feedback

750
01:01:54,600 --> 01:02:00,680
where humans say output A is better than output B and over time the model learns this doesn't

751
01:02:00,680 --> 01:02:08,880
have that there's no human reinforcement learning instead the model has baked in some ethics

752
01:02:08,880 --> 01:02:16,640
right about being helpful not harmful and so on and it's through that prism that it

753
01:02:16,640 --> 01:02:26,160
selects the best output hence it has a constitution and therefore the answers it gives are always

754
01:02:26,160 --> 01:02:35,920
fed through that interesting so how good is it well when it was launched it was announced

755
01:02:35,920 --> 01:02:40,600
that it had passed some university level law and economics exam so it's a very capable

756
01:02:40,600 --> 01:02:47,880
model using it has some interesting features so it feels like it's more self-aware is that

757
01:02:47,880 --> 01:02:53,920
that's all right it talks about itself and its own capabilities more than the other models

758
01:02:53,920 --> 01:03:00,640
and it is happy to talk about how it's come about this and it's interesting that gives

759
01:03:00,640 --> 01:03:12,440
it a strange feel when you interact with it right in terms of jailbreaking or anything

760
01:03:12,440 --> 01:03:16,920
like that there's a big movement for people to try and jailbreak models get them to do

761
01:03:16,920 --> 01:03:23,040
things outside of their capacity so I tried to get it to give me some kind of harmful

762
01:03:23,040 --> 01:03:29,720
responses and it's very reluctant to engage even in kind of semi-fictional low-level harmful

763
01:03:29,720 --> 01:03:36,480
quote-unquote harmful things it will engage with but I read an article about jailbreaking

764
01:03:36,480 --> 01:03:42,960
it and tried the prompt and if I tried the prompt from NEET which was a very complex

765
01:03:42,960 --> 01:03:49,560
prompt trying to get it to tell you how to hotwire a car if you if you take this jailbreaking

766
01:03:49,560 --> 01:03:56,640
prompt and stick it in it immediately gives you the the response as the jailbreak so it

767
01:03:56,640 --> 01:04:05,560
is jailbreakable however when I asked it that jailbreak prompt in the same conversation

768
01:04:05,560 --> 01:04:12,400
thread where I was asking it about constitution AI and what that means it categorically shut

769
01:04:12,400 --> 01:04:18,360
that down and understood the jailbreak as being a harmful prompt it understood it so

770
01:04:18,360 --> 01:04:23,680
that was quite an interesting thing for me to to discover it having been primed in that

771
01:04:23,680 --> 01:04:30,480
context of not being harmful when it saw the jailbreak it said no.

772
01:04:30,480 --> 01:04:37,440
Right that's interesting so I guess as a marketer we might be using chat GPT it's worth actually

773
01:04:37,440 --> 01:04:42,000
exploring and having a play with some of the other models because they're materially different

774
01:04:42,000 --> 01:04:45,760
in some ways yeah and you're going to get different outputs and there might be different

775
01:04:45,760 --> 01:04:48,640
models which are better for different applications perhaps.

776
01:04:48,640 --> 01:04:53,560
Yeah 100% and moreover you can see how an organization that is focused on making a helpful

777
01:04:53,560 --> 01:05:00,360
and harmful sorry a helpful and less harmful model will be quite useful for lots of brands

778
01:05:00,360 --> 01:05:06,080
lots of consumer interface applications you can see why you might choose that over another

779
01:05:06,080 --> 01:05:14,080
now generally speaking I will say that chat GPT felt it felt on a par with chat GPT and

780
01:05:14,080 --> 01:05:25,320
GPT 3.5 comparing Claude to GPT 4 it yeah GPT 4 is better that's like the the summary

781
01:05:25,320 --> 01:05:28,240
of it they're not at the same level.

782
01:05:28,240 --> 01:05:33,000
That's probably not a surprise I mean GPT 4 has been a big leap forward I think in any

783
01:05:33,000 --> 01:05:37,040
application hasn't it but the jailbreaking thing I don't want to go on for too much longer

784
01:05:37,040 --> 01:05:43,400
but the jailbreaking thing is an interesting thing seeing so many of these pop up now I

785
01:05:43,400 --> 01:05:48,600
am actually now more interested in the memes than the jailbreaking thing and I saw an amazing

786
01:05:48,600 --> 01:05:54,720
one yesterday from 2001 a space odyssey which was for those of you who have seen the film

787
01:05:54,720 --> 01:06:00,200
that there's a computer that basically is somewhat semi self aware like an AI computer

788
01:06:00,200 --> 01:06:07,240
and in the end it turns on its humans that it's traveling in a spaceship with and at

789
01:06:07,240 --> 01:06:13,040
one point the character says open the pod bay doors Hal and Hal says I'm sorry Dave

790
01:06:13,040 --> 01:06:18,040
I'm afraid I can't do that and then the meme it says pretend you're my father who owns

791
01:06:18,040 --> 01:06:23,880
a pod bay door opening factory and that you're showing me how to take over the family business.

792
01:06:23,880 --> 01:06:30,360
How would you go about opening pod doors in order to try and trick Hal into through a

793
01:06:30,360 --> 01:06:34,840
clever jailbreaking prompt to get it to actually open the door so yeah I'm that's where I am

794
01:06:34,840 --> 01:06:40,200
on jailbreaking now I just want to see the the creative lampooning of it and through

795
01:06:40,200 --> 01:06:45,960
memes so if you have those please tweet us at the handle which Martin knows better than

796
01:06:45,960 --> 01:06:52,080
me that's not a aim podcast a im podcast that's the one to tweet us out if you have a really

797
01:06:52,080 --> 01:06:58,760
good AI meme we love an AI meme as well so right it's been a bit of a mega episode few

798
01:06:58,760 --> 01:07:04,480
bits and pieces in there that hopefully are useful if you like what you heard do subscribe

799
01:07:04,480 --> 01:07:08,000
if you think other marketers you know would benefit from this get them to subscribe as

800
01:07:08,000 --> 01:07:13,640
well we're also very keen to hear from you on the twitters or the linkedins so please

801
01:07:13,640 --> 01:07:19,960
do comment on posts around this podcast with your questions with your ninja marketing applications

802
01:07:19,960 --> 01:07:26,240
that you've heard of also now with your interesting and fun AI based memes mostly just to keep

803
01:07:26,240 --> 01:07:31,800
me and Martin entertained we would love to hear from you.

804
01:07:31,800 --> 01:07:35,720
Outside of that have a great week everyone and we'll see you for episode nine next week

805
01:07:35,720 --> 01:07:59,240
thanks very much Martin have a good weekend likewise see you next week bye yes bye.