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 --> 00:00:23,000
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:23,000 --> 00:00:28,760
Welcome to Artificially Intelligent Marketing, Episode 12.

6
00:00:28,760 --> 00:00:30,600
Very exciting stuff, we're glad to have you here today.

7
00:00:30,600 --> 00:00:36,360
We're going to do our usuals, it's going to be a bunch of schnort, schnort?

8
00:00:36,360 --> 00:00:37,360
Bunch of schnort.

9
00:00:37,360 --> 00:00:41,800
We're going to have a bunch of schnort and then see where it goes from there.

10
00:00:41,800 --> 00:00:43,000
I'm looking forward to this one.

11
00:00:43,000 --> 00:00:44,000
How about it, don't.

12
00:00:44,000 --> 00:00:48,040
Now we've got a difficult decision, do we leave this in or do we?

13
00:00:48,040 --> 00:00:50,200
I think the listeners appreciate this.

14
00:00:50,200 --> 00:00:53,520
Right, I'm just going to pretend as if we cut it, even though I know you're not going

15
00:00:53,520 --> 00:00:55,000
to let me cut it.

16
00:00:55,000 --> 00:00:59,440
Welcome everyone, this week we're going to have what we always have, short snippets.

17
00:00:59,440 --> 00:01:00,440
They're going to be great.

18
00:01:00,440 --> 00:01:04,080
We're going to have a bunch of main stories as well, which is going to include some info

19
00:01:04,080 --> 00:01:08,960
about Anthropic raising a bunch of cash to build some next gen AI assistance.

20
00:01:08,960 --> 00:01:14,960
We're going to look at Microsoft's copilot coming to Windows 11, pretty exciting stuff.

21
00:01:14,960 --> 00:01:19,080
We're going to look at Google working with Europe on a stopgap AI pact.

22
00:01:19,080 --> 00:01:24,600
I swear Martin, you've written these notes to make me trip up while I try and read them.

23
00:01:24,600 --> 00:01:27,600
But I'm going to crack on regardless.

24
00:01:27,600 --> 00:01:32,760
We're going to look at the Photoshop generative AI beta launch this week, which is easy for

25
00:01:32,760 --> 00:01:37,960
folks to access if they have a subscription to Photoshop and Adobe Creative Cloud.

26
00:01:37,960 --> 00:01:44,160
And we're going to look at how Google has added generative AI to ads and product photos.

27
00:01:44,160 --> 00:01:49,360
And then finally, we're going to look at tool of the week, which is going to be Zapier.

28
00:01:49,360 --> 00:01:50,360
Right.

29
00:01:50,360 --> 00:01:53,560
Just a few bits there to get through.

30
00:01:53,560 --> 00:01:57,560
Yes, but it will all be worth it because there's loads of deep insights in there, Martin.

31
00:01:57,560 --> 00:02:00,840
So I think it's going to be a blooming good time.

32
00:02:00,840 --> 00:02:04,280
Let's start with those short snippets first, shall we?

33
00:02:04,280 --> 00:02:10,880
The first one is that news actually today, we're recording this on Friday, 26th of May,

34
00:02:10,880 --> 00:02:17,000
is that OpenAI could leave the EU if the new AI rules, which we've been talking about before

35
00:02:17,000 --> 00:02:20,240
on the podcast, are passed.

36
00:02:20,240 --> 00:02:26,200
Sam Altman, who is CEO of OpenAI said, either we'll be able to solve for these requirements

37
00:02:26,200 --> 00:02:27,440
or not.

38
00:02:27,440 --> 00:02:28,720
If we can comply, we will.

39
00:02:28,720 --> 00:02:31,080
And if we can't, well, we'll cease operating.

40
00:02:31,080 --> 00:02:34,800
We will try, but there are technical limits to what's possible.

41
00:02:34,800 --> 00:02:42,640
That could be quite interesting for users of tools that require OpenAI as the element

42
00:02:42,640 --> 00:02:46,960
that underpins the tools, or if indeed you're using ChatGPT itself and you're based in the

43
00:02:46,960 --> 00:02:47,960
EU.

44
00:02:47,960 --> 00:02:54,440
This could open up opportunities though, if other, if certainly OpenAI and maybe other

45
00:02:54,440 --> 00:03:00,360
US based firms could open up opportunities for AI firms based in Europe and that are

46
00:03:00,360 --> 00:03:06,280
US, sorry, EU compliant to gain some market traction by dominating in the EU first.

47
00:03:06,280 --> 00:03:10,840
So I doubt we'd get that far, if I'm honest, but it'd be interesting to see how this opens

48
00:03:10,840 --> 00:03:12,880
up.

49
00:03:12,880 --> 00:03:17,880
Next short snippet is A16Z released its AI cannon this week.

50
00:03:17,880 --> 00:03:21,280
You probably saw some posts, there's plenty of posts on social about this.

51
00:03:21,280 --> 00:03:27,880
It is a mega AI resource with loads of information about how AI works, the different types of

52
00:03:27,880 --> 00:03:31,000
models, there's market analysis in there.

53
00:03:31,000 --> 00:03:35,600
It is a real one stop location for all things AI.

54
00:03:35,600 --> 00:03:40,240
So if until this point, haven't really been following much of what's going on, but you're

55
00:03:40,240 --> 00:03:43,160
like crumbs, I really need to do a deep dive.

56
00:03:43,160 --> 00:03:47,680
Go to that webpage, it's got everything you need and we'll pop it in the show notes, won't

57
00:03:47,680 --> 00:03:48,680
we Martin?

58
00:03:48,680 --> 00:03:53,480
Yeah, it's got some fantastic sections in it from a nice gentle introduction through

59
00:03:53,480 --> 00:03:58,360
to the kind of foundational learning, even kind of going to the real fundamentals.

60
00:03:58,360 --> 00:04:03,640
I'd recommend people check out the Word to VEC explainer, which is the absolute basic

61
00:04:03,640 --> 00:04:06,320
building block of large language models.

62
00:04:06,320 --> 00:04:09,520
I think if you can get your head around that, then you're a long way to understanding how

63
00:04:09,520 --> 00:04:10,520
these things work.

64
00:04:10,520 --> 00:04:12,320
So yeah, great resource.

65
00:04:12,320 --> 00:04:13,320
Agreed.

66
00:04:13,320 --> 00:04:15,400
Spend Friday night reading all of that.

67
00:04:15,400 --> 00:04:18,800
Spend a night impress your friends at the pub.

68
00:04:18,800 --> 00:04:20,600
I think that will be both.

69
00:04:20,600 --> 00:04:27,840
If there's one thing I've learned in recent weeks, people love hearing someone more on

70
00:04:27,840 --> 00:04:30,160
about AI in a pub.

71
00:04:30,160 --> 00:04:34,280
But no, but listen, vector databases, listen.

72
00:04:34,280 --> 00:04:35,280
Vector embeddings.

73
00:04:35,280 --> 00:04:36,280
Yes.

74
00:04:36,280 --> 00:04:43,080
The way that cat and dog are close to each other in a way that cat and house are not.

75
00:04:43,080 --> 00:04:46,600
I'm pretty sure this was part of the original version of how to win friends and influence

76
00:04:46,600 --> 00:04:49,800
people and of course it wasn't, but maybe we should update it.

77
00:04:49,800 --> 00:04:53,200
We could probably get AI to update it, but that's a different topic we're discussing

78
00:04:53,200 --> 00:04:54,200
in the pub.

79
00:04:54,200 --> 00:04:56,640
Apologies listeners for the digression.

80
00:04:56,640 --> 00:04:58,760
Back into the short snippets.

81
00:04:58,760 --> 00:05:03,360
Nvidia's stock price went boom shakalaka after greater than expected earnings.

82
00:05:03,360 --> 00:05:10,200
Presumably everyone and their nan are buying GPUs to run generative AI models on and that's

83
00:05:10,200 --> 00:05:12,400
what's boosted that.

84
00:05:12,400 --> 00:05:17,200
We've also previously reported on how many large players are investing in developing

85
00:05:17,200 --> 00:05:19,200
their own chips.

86
00:05:19,200 --> 00:05:24,600
So whether or not this progresses and Nvidia's stock price continues to go shooting up because

87
00:05:24,600 --> 00:05:28,880
their revenues go shooting up or whether most of these companies end up developing their

88
00:05:28,880 --> 00:05:29,880
own chips, who knows?

89
00:05:29,880 --> 00:05:34,400
But for now, if you own stock in Nvidia, you're doing rather nicely.

90
00:05:34,400 --> 00:05:38,160
We don't, we're not experts in the financial markets and we don't offer any investment

91
00:05:38,160 --> 00:05:39,160
advice.

92
00:05:39,160 --> 00:05:42,320
A quick disclaimer there.

93
00:05:42,320 --> 00:05:48,200
Next bit of news is free users of ChatGPT can now access the ability to browse the web

94
00:05:48,200 --> 00:05:50,880
or use the web browser plugin, I should say.

95
00:05:50,880 --> 00:05:55,920
So even if you're a free user, now it will actually access up to date content through

96
00:05:55,920 --> 00:05:56,920
that plugin.

97
00:05:56,920 --> 00:05:59,720
It's powered by and branded as Bing.

98
00:05:59,720 --> 00:06:05,760
So I guess that's part of the trade off there in terms of being able to get to access it.

99
00:06:05,760 --> 00:06:07,800
And then you were looking at a story about water.

100
00:06:07,800 --> 00:06:09,600
What's this water story, Martin, tell us.

101
00:06:09,600 --> 00:06:13,640
Yeah, so it turns out there's quite a bit of energy and water consumption involved in

102
00:06:13,640 --> 00:06:15,520
using these large language models.

103
00:06:15,520 --> 00:06:19,680
There was a piece of research published this week, we'll link you to the paper.

104
00:06:19,680 --> 00:06:28,080
It says ChatGPT may have used 255 million liters of water in a single month.

105
00:06:28,080 --> 00:06:33,840
Yeah, so it has not only massive energy consumption, but the water used to call the data centers

106
00:06:33,840 --> 00:06:37,200
is also enormous.

107
00:06:37,200 --> 00:06:45,520
So in April, ChatGPT reportedly had 1.7 billion visitors who each made an average of six queries.

108
00:06:45,520 --> 00:06:52,240
So that's a total of 10 billion, just over 10 billion queries in April.

109
00:06:52,240 --> 00:07:03,360
And it's thought that it consumes 0.025 liters of water per query, which gives you that 255

110
00:07:03,360 --> 00:07:04,880
million liters of water.

111
00:07:04,880 --> 00:07:10,880
And for comparison, the average American family uses 40,000 liters in a month.

112
00:07:10,880 --> 00:07:13,800
So it's quite a lot.

113
00:07:13,800 --> 00:07:18,720
Interestingly I was at a workshop while I was delivering a workshop on using ChatGPT

114
00:07:18,720 --> 00:07:23,000
this week and mentioned this figure.

115
00:07:23,000 --> 00:07:29,960
And in the room there was, I would call it at best, a kind of disinterested shrug.

116
00:07:29,960 --> 00:07:30,960
Ambivalence.

117
00:07:30,960 --> 00:07:33,540
Absolute ambivalence, yeah, totally.

118
00:07:33,540 --> 00:07:35,960
People just felt that...

119
00:07:35,960 --> 00:07:42,320
And people actually started getting into the technical discussion about how watering data

120
00:07:42,320 --> 00:07:44,400
centers is probably on a closed loop system.

121
00:07:44,400 --> 00:07:45,400
So what's the problem?

122
00:07:45,400 --> 00:07:47,120
And I backed out at that point.

123
00:07:47,120 --> 00:07:48,120
I had no idea.

124
00:07:48,120 --> 00:07:52,480
That's above my pay grade or knowledge, I have to say.

125
00:07:52,480 --> 00:07:57,200
One thing I do remember a while back is having a discussion with a potential client, a buyer

126
00:07:57,200 --> 00:08:04,400
straight back in the day who had data servers and stuff all based in Northern Iceland, where

127
00:08:04,400 --> 00:08:06,640
it's proper cold most of the time.

128
00:08:06,640 --> 00:08:12,160
It seemed like quite a clever way to keep your massive server farms cold and pulling

129
00:08:12,160 --> 00:08:13,360
in cold air from outside.

130
00:08:13,360 --> 00:08:14,960
Which is pretty cool.

131
00:08:14,960 --> 00:08:16,480
Anyway, another digression.

132
00:08:16,480 --> 00:08:18,660
It's already been an episode of digressions.

133
00:08:18,660 --> 00:08:21,320
People just hitting stop, pause, unsubscribe.

134
00:08:21,320 --> 00:08:22,320
Don't do it.

135
00:08:22,320 --> 00:08:26,640
I promise there's fantastic insights, even greater insights than what you've had already.

136
00:08:26,640 --> 00:08:28,920
Actually coming after this.

137
00:08:28,920 --> 00:08:31,680
And actually we're going to get there now because we're going to launch into our first

138
00:08:31,680 --> 00:08:39,440
big story of the week, which is Anthropic raising $450 million to build next gen AI

139
00:08:39,440 --> 00:08:40,440
assistance.

140
00:08:40,440 --> 00:08:42,280
Martin, you spied this.

141
00:08:42,280 --> 00:08:45,960
Tell us what you found and why it's interesting for marketers.

142
00:08:45,960 --> 00:08:47,720
We've spoken about Claude in recent weeks.

143
00:08:47,720 --> 00:08:50,480
This is a continuation of our interest in this story.

144
00:08:50,480 --> 00:08:55,280
So Anthropic are doing a really good job of raising their profile and fundraising.

145
00:08:55,280 --> 00:09:01,200
So 450 million raised in series C funding.

146
00:09:01,200 --> 00:09:08,560
It's thought that it values Anthropic at over 4.1 billion and there are some interesting

147
00:09:08,560 --> 00:09:10,440
venture groups backing them.

148
00:09:10,440 --> 00:09:15,860
So Google, Salesforce and Zoom are all in on this round.

149
00:09:15,860 --> 00:09:22,000
So Anthropic is aiming to build next gen AI systems focused principally on safety, reliability

150
00:09:22,000 --> 00:09:29,320
and honesty, which is through their Claude Bay system, which is a constitutional AI designed

151
00:09:29,320 --> 00:09:34,920
to be helpful, not harmful and honest.

152
00:09:34,920 --> 00:09:41,200
So another interesting point that came from this is that Zoom have also announced a partnership

153
00:09:41,200 --> 00:09:47,240
with Anthropic to build customer facing AI products principally focused on reliability,

154
00:09:47,240 --> 00:09:54,200
productivity and safety following a similar tie up with Google.

155
00:09:54,200 --> 00:09:59,040
All of this gives Anthropic a good position in the market in terms of being able to compete

156
00:09:59,040 --> 00:10:00,440
against the likes of OpenAI.

157
00:10:00,440 --> 00:10:03,520
They're clearly getting access to some big players.

158
00:10:03,520 --> 00:10:09,240
The models are going to be finding some interesting use cases amongst tech stacks that we all

159
00:10:09,240 --> 00:10:12,760
know and use every day.

160
00:10:12,760 --> 00:10:16,760
The real differentiator for Anthropic at the moment is, as we have mentioned previously,

161
00:10:16,760 --> 00:10:20,480
that 100,000 token context window.

162
00:10:20,480 --> 00:10:27,560
So your prompts can be properly long, like novel length long.

163
00:10:27,560 --> 00:10:32,480
Apparently according to the story, Anthropic plans to build an even larger next gen model,

164
00:10:32,480 --> 00:10:43,880
which they say requires 10 to the 25 flops of power.

165
00:10:43,880 --> 00:10:50,880
So that will require one billion pound funding in the next 18 months.

166
00:10:50,880 --> 00:10:54,560
So that's not a cheap endeavor.

167
00:10:54,560 --> 00:11:00,280
Apparently hoping to raise five billion over the next two years to expand its products.

168
00:11:00,280 --> 00:11:04,680
So really they're properly going up against it.

169
00:11:04,680 --> 00:11:12,720
Anthropocare are getting deep pockets in order to be able to compete with OpenAI, who themselves

170
00:11:12,720 --> 00:11:15,640
had 10 billion pound of investment from Microsoft.

171
00:11:15,640 --> 00:11:20,720
Yeah, it's an interesting one because I think it shows that as marketers, we're going to

172
00:11:20,720 --> 00:11:24,800
probably end up with tools as we've talked about on the podcast before, that are driven

173
00:11:24,800 --> 00:11:30,080
by different large language models, which is probably good given that it hasn't really

174
00:11:30,080 --> 00:11:33,680
happened for a while, but in the early days of chat GPT, when it went down a few times

175
00:11:33,680 --> 00:11:40,240
and people were like, oh, I got a thing for myself, which was obviously extremely challenging

176
00:11:40,240 --> 00:11:44,080
for the couple of hours that it was down for people who were heavy users.

177
00:11:44,080 --> 00:11:50,800
But we've also touched on how different models appear to have different strengths.

178
00:11:50,800 --> 00:11:56,080
So actually seeing that we're going to have this slightly more diverse ecosystem than

179
00:11:56,080 --> 00:12:00,540
perhaps we might have imagined when it was looking in the early days, like it was already

180
00:12:00,540 --> 00:12:06,040
an OpenAI takes all game, probably a good thing for us in terms of diversity of tools

181
00:12:06,040 --> 00:12:11,080
that can do different stuff for different applications, strengths and weaknesses.

182
00:12:11,080 --> 00:12:12,080
Most definitely.

183
00:12:12,080 --> 00:12:17,640
And having a system that is fundamentally designed to be different from something like

184
00:12:17,640 --> 00:12:24,600
GPT-4, which is the RLHF model, the reinforcement learning through human feedback model, having

185
00:12:24,600 --> 00:12:28,120
that constitutional framework gives it a real differentiator.

186
00:12:28,120 --> 00:12:35,040
And it's nice to see these on the market because a lot of the models are RLHF.

187
00:12:35,040 --> 00:12:37,560
So yeah, it's a good look to Anthropic.

188
00:12:37,560 --> 00:12:40,120
I'm really enjoying using Claude at the moment.

189
00:12:40,120 --> 00:12:49,640
It's probably my go-to model, even though chat GPT at the moment is expanding with lots

190
00:12:49,640 --> 00:12:51,400
of the plugins.

191
00:12:51,400 --> 00:12:57,480
The real-time connected nature to the web is interesting, of course, being able to browse

192
00:12:57,480 --> 00:13:03,200
the web with GPT-4 is cool, but it's incredibly slow.

193
00:13:03,200 --> 00:13:07,520
I actually find that I can go online, find an article, copy and paste it, stick it into

194
00:13:07,520 --> 00:13:10,560
Claude and say, give me the key facts quicker.

195
00:13:10,560 --> 00:13:14,320
And I don't have to worry about character length because, as I've mentioned already,

196
00:13:14,320 --> 00:13:17,040
you can stick a novel in there.

197
00:13:17,040 --> 00:13:19,080
Yeah.

198
00:13:19,080 --> 00:13:21,640
My jealous green eyes are gleaming now.

199
00:13:21,640 --> 00:13:26,240
And in fact, with the variety of easy image manipulation tools available now, I could

200
00:13:26,240 --> 00:13:29,760
probably take a screen grab and mark it up pretty quickly and easily at this point, but

201
00:13:29,760 --> 00:13:32,160
get onto Photoshop later.

202
00:13:32,160 --> 00:13:37,520
I did see a Twitter, I think it was on Twitter, it might have been LinkedIn, conversation

203
00:13:37,520 --> 00:13:45,120
about how the real power of generative AI might turn out to be content summarization,

204
00:13:45,120 --> 00:13:48,280
not content generation, which I think is really interesting.

205
00:13:48,280 --> 00:13:54,400
We've talked previously about this concept that someone might write a long memo or email

206
00:13:54,400 --> 00:13:59,240
that then gets summarized down by an AI for someone to quickly absorb.

207
00:13:59,240 --> 00:14:04,960
They maybe dictate or write a fairly long response that then gets summarized down.

208
00:14:04,960 --> 00:14:10,520
And actually the real gatekeepers of what information makes it through and what doesn't

209
00:14:10,520 --> 00:14:13,240
might actually be the AI summarizing tools.

210
00:14:13,240 --> 00:14:17,800
But I am jealous of you having access to Claude for that summarization reason because it does

211
00:14:17,800 --> 00:14:19,760
look pretty powerful for that.

212
00:14:19,760 --> 00:14:27,640
On that point about summarization, if anybody hasn't read Wired's article about how they

213
00:14:27,640 --> 00:14:33,040
plan to use generative AI, I highly recommend they check it out.

214
00:14:33,040 --> 00:14:34,120
They published it a few months ago.

215
00:14:34,120 --> 00:14:37,120
I think we might have even mentioned it on a previous episode.

216
00:14:37,120 --> 00:14:41,820
But it mentions in there about they will not use generative AI or large language models

217
00:14:41,820 --> 00:14:43,840
to do any editing.

218
00:14:43,840 --> 00:14:48,880
And they say because it's always a judgment call and what stays and what goes, what is

219
00:14:48,880 --> 00:14:52,160
important and what isn't is going to be different for different people.

220
00:14:52,160 --> 00:14:58,440
So a robot is just a different agent in that aspect.

221
00:14:58,440 --> 00:15:01,080
It's always going to be left to a human to make those calls.

222
00:15:01,080 --> 00:15:02,080
Interesting.

223
00:15:02,080 --> 00:15:03,080
Yeah, I agree.

224
00:15:03,080 --> 00:15:08,800
And I think we may have even talked previously about the dangers of letting a single AI agent

225
00:15:08,800 --> 00:15:13,720
make the call on what gets into the summary and what doesn't.

226
00:15:13,720 --> 00:15:16,600
I think we talked previously about you and I read a book, Martin, and we're going to

227
00:15:16,600 --> 00:15:21,680
take different things away from it because of our context, our experience, our background,

228
00:15:21,680 --> 00:15:25,960
our particular interests, why we read the book, what particular sections we focused

229
00:15:25,960 --> 00:15:26,960
on the most.

230
00:15:26,960 --> 00:15:32,560
And yes, but certainly it's good to have all those different tools.

231
00:15:32,560 --> 00:15:37,600
And I think it's going to be a watch this space as the Anthropic continues to expand

232
00:15:37,600 --> 00:15:40,360
and evolve its offering in this market.

233
00:15:40,360 --> 00:15:48,080
Right, next, let's talk about Microsoft's announcing Windows Copilot, an AI personal

234
00:15:48,080 --> 00:15:51,040
assistant for Windows 11.

235
00:15:51,040 --> 00:15:56,960
So another previous episode callback here, but Martin, we've talked previously about

236
00:15:56,960 --> 00:16:01,040
the edge sidebar and making the transition from Chrome to Edge.

237
00:16:01,040 --> 00:16:02,880
And honestly, I wouldn't look back at this point.

238
00:16:02,880 --> 00:16:03,880
I'm really glad I did it.

239
00:16:03,880 --> 00:16:09,620
I'm absolutely in and out of that sidebar, having Bing available to push content into

240
00:16:09,620 --> 00:16:14,200
to quickly summarize or to more widely search is actually proving really useful.

241
00:16:14,200 --> 00:16:18,500
But of course, Windows and Microsoft, they want to take that a bit further.

242
00:16:18,500 --> 00:16:25,000
So with Windows Copilot, Windows 11 will effectively have a centralized AI assistant that you can

243
00:16:25,000 --> 00:16:27,560
tap into at any point.

244
00:16:27,560 --> 00:16:35,640
So it's the equivalent of the Bing sidebar in Edge, but having that basically in every

245
00:16:35,640 --> 00:16:41,840
app that you're using across Windows and actually basically having it as you're using Windows,

246
00:16:41,840 --> 00:16:43,840
whichever platforms you're working in.

247
00:16:43,840 --> 00:16:50,880
So that in itself is pretty interesting and powerful because we've talked about the where

248
00:16:50,880 --> 00:16:55,760
are the moats for different companies and which products should a market to buy and

249
00:16:55,760 --> 00:16:59,440
invest their time and money and training resource and energy in.

250
00:16:59,440 --> 00:17:07,320
And ultimately, because so many users already use Google Apps or Microsoft Office, there's

251
00:17:07,320 --> 00:17:11,360
an argument to say, wait until this power comes to those because your team is already

252
00:17:11,360 --> 00:17:14,320
using them, etc, etc, etc.

253
00:17:14,320 --> 00:17:18,080
This is an area where Microsoft even has one up on Google, right?

254
00:17:18,080 --> 00:17:23,080
Because so many people use Windows as their operating environment.

255
00:17:23,080 --> 00:17:27,800
And yes, I appreciate Chromebooks and stuff running on Android and also mobile devices.

256
00:17:27,800 --> 00:17:32,440
But for most people in the work environment, still using laptops of some variety, appreciate

257
00:17:32,440 --> 00:17:34,840
Mac users out there as well.

258
00:17:34,840 --> 00:17:40,640
But yes, having that power baked into not just Microsoft Office, but Windows itself

259
00:17:40,640 --> 00:17:42,800
is pretty interesting.

260
00:17:42,800 --> 00:17:50,280
There's also the sort of co-announcement that goes along with this is extending chat plugins

261
00:17:50,280 --> 00:17:51,560
to work in Windows.

262
00:17:51,560 --> 00:17:57,360
So in essence, you can start to imagine this plugin app ecosystem that we talked about

263
00:17:57,360 --> 00:18:03,840
previously for chat GPT opening up so that people can augment your chat experience within

264
00:18:03,840 --> 00:18:07,280
Windows with a whole host of plugins.

265
00:18:07,280 --> 00:18:12,440
So potentially it could become very powerful tool indeed.

266
00:18:12,440 --> 00:18:19,160
And maybe even another play by Microsoft to see if they can control which tool do you

267
00:18:19,160 --> 00:18:21,320
turn to first, right?

268
00:18:21,320 --> 00:18:27,700
Like if you just got your co-pilot sidebar always open in Windows, well then every search

269
00:18:27,700 --> 00:18:34,760
that you do of your own files, of the web, of your own data, maybe generative text creation,

270
00:18:34,760 --> 00:18:37,520
image creation might all start in there.

271
00:18:37,520 --> 00:18:41,840
And of course, if they control that, then they control the ecosystem one, assume.

272
00:18:41,840 --> 00:18:46,960
So pretty interesting stuff here and be interested in see what that looks like when it starts

273
00:18:46,960 --> 00:18:49,000
to roll out.

274
00:18:49,000 --> 00:18:53,800
I demoed this to the group that I was talking about yesterday, or I showed them the video

275
00:18:53,800 --> 00:18:56,880
trailer from the announcement.

276
00:18:56,880 --> 00:18:59,680
And this comes off the back of having done a previous session with the same group where

277
00:18:59,680 --> 00:19:03,480
we'd looked at the co-pilot demo as well.

278
00:19:03,480 --> 00:19:09,760
And one of the questions from people in the room was, which one do we go to for what?

279
00:19:09,760 --> 00:19:12,400
And do they talk to each other?

280
00:19:12,400 --> 00:19:19,120
Is Windows co-pilot going to be the same as the Office co-pilot?

281
00:19:19,120 --> 00:19:25,320
Because the demo talks about being able to interrogate your files, you can drag and drop

282
00:19:25,320 --> 00:19:30,480
a PDF into it and then take content from that and put it into a different document format.

283
00:19:30,480 --> 00:19:35,800
So I think making sure that they get this right and it isn't a jarring, confusing experience

284
00:19:35,800 --> 00:19:42,840
where co-pilot for Windows can do one thing and Office can do another thing.

285
00:19:42,840 --> 00:19:48,640
I think they need to get almost like a Microsoft co-pilot the end.

286
00:19:48,640 --> 00:19:53,440
This is one assistant that works across the suite of tools.

287
00:19:53,440 --> 00:20:00,000
Because at the moment, the demo makes it look like, well, we've got GitHub co-pilot and

288
00:20:00,000 --> 00:20:09,760
we've got Windows co-pilot and we've got Office 365 co-pilot and I just hope they make it

289
00:20:09,760 --> 00:20:11,680
a uniform experience.

290
00:20:11,680 --> 00:20:16,000
You've said previously that because a lot of these large language models have similar

291
00:20:16,000 --> 00:20:20,500
strengths and weaknesses, that this could end up being a battle of UX, right?

292
00:20:20,500 --> 00:20:25,080
And having a high quality user experience and solving for the problem you just mentioned

293
00:20:25,080 --> 00:20:26,520
is a big part of that.

294
00:20:26,520 --> 00:20:33,640
I very much hope it's a sidebar that takes up a little bit of screen real estate at all

295
00:20:33,640 --> 00:20:39,520
times honestly, whereby it knows what window I have open.

296
00:20:39,520 --> 00:20:40,520
It knows I'm in...

297
00:20:40,520 --> 00:20:41,520
Context aware.

298
00:20:41,520 --> 00:20:42,520
Yeah, it's context aware.

299
00:20:42,520 --> 00:20:45,720
It knows that I'm in Word and so there's certain things it's going to do in Word.

300
00:20:45,720 --> 00:20:49,920
And actually if I'm not in Word and I'm like, oh, I need you to summarize this in a document,

301
00:20:49,920 --> 00:20:54,880
then it just does it, pushes it straight into a Word doc and then it's context aware enough

302
00:20:54,880 --> 00:20:58,200
to say, hey, what else do you want to do with this?

303
00:20:58,200 --> 00:21:01,840
Now we're in Word together because I know as well as you know, Paul.

304
00:21:01,840 --> 00:21:06,600
Crumbs, could you imagine the conversations I'm going to have with AI, knowing some of

305
00:21:06,600 --> 00:21:10,080
my communication limitations.

306
00:21:10,080 --> 00:21:12,440
So yeah, I'm in agreement with you.

307
00:21:12,440 --> 00:21:19,920
I think that could be the thing that makes it a winner or a loser because if it's hard

308
00:21:19,920 --> 00:21:23,360
to use or confusing, people are going to get annoyed, aren't they?

309
00:21:23,360 --> 00:21:24,360
They will.

310
00:21:24,360 --> 00:21:30,760
So a general observation about these AI assistants at the kind of system level or both the OS

311
00:21:30,760 --> 00:21:37,120
and I guess at the kind of suite level, so Office 365 Copilot as well.

312
00:21:37,120 --> 00:21:44,840
They talked about something in the Office 365 Copilot demo where they said most users

313
00:21:44,840 --> 00:21:51,520
of PowerPoint are using 10% of the functionality that it actually has.

314
00:21:51,520 --> 00:21:56,320
And using Copilot will enable people to unlock all of this additional functionality that

315
00:21:56,320 --> 00:21:58,320
is baked into the software.

316
00:21:58,320 --> 00:22:02,040
And I think that's probably the case with operating systems as well.

317
00:22:02,040 --> 00:22:04,640
You get into a habit, you know how to do certain things.

318
00:22:04,640 --> 00:22:08,120
If you're not somebody who is confident in a PC, so you sit down, you turn it on, you

319
00:22:08,120 --> 00:22:14,200
do your work and you would never dare venture into the control panel even, this is going

320
00:22:14,200 --> 00:22:19,280
to make your interactions with your computer much easier, one would hope.

321
00:22:19,280 --> 00:22:25,280
So I think that's going to be positive because we sit at these machines all the time and

322
00:22:25,280 --> 00:22:26,900
barely scratch the surface.

323
00:22:26,900 --> 00:22:34,480
We get into routines of doing what we know and if we can start to get the full potential

324
00:22:34,480 --> 00:22:38,480
of the tools that we have at our disposal already out of it through these assistants

325
00:22:38,480 --> 00:22:39,760
then yay.

326
00:22:39,760 --> 00:22:42,800
I completely agree.

327
00:22:42,800 --> 00:22:51,160
When we talked about the code interpreter plugin for Chatch GPT and its ability to analyze

328
00:22:51,160 --> 00:22:57,440
data, uncover trends, report on those trends, write paper abstracts and things like that,

329
00:22:57,440 --> 00:23:02,840
we talked about how it's critical that you have an informed human there to know what

330
00:23:02,840 --> 00:23:07,280
is an interesting trend or to ask great questions of the data.

331
00:23:07,280 --> 00:23:12,000
But what I would really love is what you described and then up to the next level again, which

332
00:23:12,000 --> 00:23:18,840
is when I'm in Excel that the Co-Pilot tool has enough insights into what I'm doing and

333
00:23:18,840 --> 00:23:25,720
how I work to actually be able to say, hey do you know that thing you're doing, I could

334
00:23:25,720 --> 00:23:31,240
build you a macro to do that, you don't have to do it by brute force and just manually

335
00:23:31,240 --> 00:23:35,200
like what you are and then I go, oh that sounds amazing and we go yeah no worries I'll create

336
00:23:35,200 --> 00:23:38,400
the macro for you and all you've got to do is press that red button at the top, well

337
00:23:38,400 --> 00:23:42,800
maybe not it'll be like oh I've started the macro, tell me what next CSV you want to feed

338
00:23:42,800 --> 00:23:48,120
into it or whatever because that would be the cool part because the extra part of that

339
00:23:48,120 --> 00:23:53,000
is you don't know what you don't know sometimes do in terms of you don't use that 90% sometimes

340
00:23:53,000 --> 00:23:58,400
because it's hard to technically do but sometimes because you don't even know it can do it.

341
00:23:58,400 --> 00:24:01,880
Yeah absolutely, particularly with the Microsoft suite.

342
00:24:01,880 --> 00:24:05,280
It's actually why I use Google Workspace a lot of the time because I always think Google

343
00:24:05,280 --> 00:24:12,520
Workspace all of their suite of tools they're not as capable as the Office suite but they're

344
00:24:12,520 --> 00:24:15,600
more than capable enough for what I typically need.

345
00:24:15,600 --> 00:24:20,520
Agreed right we better move on to the next story before we run out of time Martin so

346
00:24:20,520 --> 00:24:27,940
let's go to Google to work with you upon the Stopgap AI pact, thanks again for that.

347
00:24:27,940 --> 00:24:30,360
Tell us Martin what you saw here this week.

348
00:24:30,360 --> 00:24:35,640
So a bit of a follow-up to the story that we've been following over a number of weeks

349
00:24:35,640 --> 00:24:41,680
which is the AI act coming in from the EU and following on from the Sam Altman comments

350
00:24:41,680 --> 00:24:45,800
in the short snippets easy for me to say.

351
00:24:45,800 --> 00:24:51,240
So this story basically what's happened is the EU Commission or the European Commission

352
00:24:51,240 --> 00:24:56,480
and Google have come together and there will be some sort of Stopgap AI pact as you mentioned

353
00:24:56,480 --> 00:25:02,040
in the headline but the story will involve all major companies working on AI both in

354
00:25:02,040 --> 00:25:06,760
Europe and outside of Europe that is the goal at least because the AI act is due to come

355
00:25:06,760 --> 00:25:12,280
in in two years time from sign off or implementation.

356
00:25:12,280 --> 00:25:19,560
There's going to be this two year period where AI companies are not going to have to do anything

357
00:25:19,560 --> 00:25:23,560
and regulators are a bit like oh actually the pace of change in this space is so rapid

358
00:25:23,560 --> 00:25:28,240
that two years with nothing is a bit scary.

359
00:25:28,240 --> 00:25:33,960
So the goal of this AI pact is to mitigate the gravest risks associated with this rapidly

360
00:25:33,960 --> 00:25:38,240
evolving technology until proper legislation is put in place.

361
00:25:38,240 --> 00:25:44,360
EU Commissioner Thierry Breton said Sundar Pichai and I have agreed that we cannot afford

362
00:25:44,360 --> 00:25:50,040
to wait until AI regulation actually becomes applicable and to work together with all AI

363
00:25:50,040 --> 00:25:57,360
partners to develop an AI pact on a voluntary basis ahead of the legal deadline.

364
00:25:57,360 --> 00:26:03,320
I've already I already have a common vision of what could be put in place in anticipation

365
00:26:03,320 --> 00:26:08,560
and which could allow us to give some elements of protection which suggests that this pact

366
00:26:08,560 --> 00:26:13,920
will aim to establish some basic rules or guidelines around AI.

367
00:26:13,920 --> 00:26:22,120
However this will be voluntary so this will not be enforceable and this is going to be

368
00:26:22,120 --> 00:26:30,840
pretty much a good faith deal between the likes of Google, Microsoft, OpenAI and the

369
00:26:30,840 --> 00:26:31,840
EU.

370
00:26:31,840 --> 00:26:39,160
But it should help establish or I think the goal of this is maybe that it helps to establish

371
00:26:39,160 --> 00:26:44,440
a framework for responsible AI development in the future.

372
00:26:44,440 --> 00:26:47,600
In the coverage of this story not much detail is given.

373
00:26:47,600 --> 00:26:51,480
When I say not much no detail is given.

374
00:26:51,480 --> 00:26:53,440
Zero!

375
00:26:53,440 --> 00:26:58,000
Absolutely no detail at all and this is covered by Reuters and a bunch of other reputable

376
00:26:58,000 --> 00:26:59,000
sources.

377
00:26:59,000 --> 00:27:05,040
So we don't know what this pact involves but Breton has mentioned the possibility of

378
00:27:05,040 --> 00:27:11,960
labelling AI systems which would probably be some sort of categorisation based on levels

379
00:27:11,960 --> 00:27:18,440
of risk, transparency and other attributes that we already know are important and are

380
00:27:18,440 --> 00:27:20,880
kind of covered in the AI act.

381
00:27:20,880 --> 00:27:28,400
The AI act is going to be a very risk-based human centric piece of legislation and regulation

382
00:27:28,400 --> 00:27:33,680
and so it would make sense that they're having those conversations in anticipation of the

383
00:27:33,680 --> 00:27:36,080
legislation coming in down the line.

384
00:27:36,080 --> 00:27:42,520
I also read something this week, I can't remember the source for it but it suggested that people

385
00:27:42,520 --> 00:27:50,160
wouldn't be surprised if the AI act gets watered down slightly in terms of requirements for

386
00:27:50,160 --> 00:27:57,000
transparency because explainability at the moment is still such a technically challenging

387
00:27:57,000 --> 00:27:58,000
problem.

388
00:27:58,000 --> 00:28:05,000
I think this comes to the point that Sam Altman said in terms of leaving the EU that if the

389
00:28:05,000 --> 00:28:10,880
requirements were such that they just couldn't meet them due to technical reasons then they

390
00:28:10,880 --> 00:28:13,120
would leave.

391
00:28:13,120 --> 00:28:17,600
As you said when you were reporting on that Paul, you don't think it would come to that.

392
00:28:17,600 --> 00:28:20,200
I am not entirely sure it would either.

393
00:28:20,200 --> 00:28:24,840
I think the regulators will work.

394
00:28:24,840 --> 00:28:30,680
Fragmatically, there'll be more focus on, or in my opinion, the EU will be more focused

395
00:28:30,680 --> 00:28:39,440
on the rights of individuals rather than concerns related to copyright or something like that.

396
00:28:39,440 --> 00:28:45,360
Yeah, it'd be interesting to see that play out and neither of us are lawyers or legal

397
00:28:45,360 --> 00:28:46,360
professionals.

398
00:28:46,360 --> 00:28:49,760
Lots of caveats in today's episode.

399
00:28:49,760 --> 00:28:52,160
But yeah, I think you're right.

400
00:28:52,160 --> 00:28:58,680
When you frame it like that and you look at how Sam Altman phrased his comments, it sounds

401
00:28:58,680 --> 00:29:03,920
like they put more than a reasonable amount of effort to try and meet whatever the stipulations

402
00:29:03,920 --> 00:29:04,920
were.

403
00:29:04,920 --> 00:29:10,320
One would have to assume that any other producer of large language models might also struggle

404
00:29:10,320 --> 00:29:15,200
to meet those stipulations on transparency and explainability.

405
00:29:15,200 --> 00:29:21,280
I think maybe the critical one for OpenAI is they didn't really tell us much about GPT

406
00:29:21,280 --> 00:29:27,280
course training dataset and needing to be able to provide full access and insight into

407
00:29:27,280 --> 00:29:32,120
what that dataset was might be one of the kickers for them at least.

408
00:29:32,120 --> 00:29:39,440
Yeah, and you would think that this could be overcome by having some sort of auditing

409
00:29:39,440 --> 00:29:49,560
mechanism which allows for commercial sensitivities where they are scrutinizable by those that

410
00:29:49,560 --> 00:29:57,040
need to scrutinize the public bodies but not necessarily made publicly available for all

411
00:29:57,040 --> 00:29:59,000
and sundry to see and interrogate.

412
00:29:59,000 --> 00:30:07,760
That doesn't seem like it's beyond the capabilities of man to figure this out.

413
00:30:07,760 --> 00:30:13,600
No, and not our responsibility either, thankfully, Martin.

414
00:30:13,600 --> 00:30:17,680
But yeah, I think it'd be interesting for marketers to keep an eye on how all of this

415
00:30:17,680 --> 00:30:23,720
plays out because again, as we are all starting to think about how we can leverage AI and

416
00:30:23,720 --> 00:30:28,760
different aspects of our business and our marketing processes to increase creativity,

417
00:30:28,760 --> 00:30:36,000
improve efficiency, then understanding what types of frameworks are going to limit what

418
00:30:36,000 --> 00:30:40,240
types of tools before you make investments in building this into your business.

419
00:30:40,240 --> 00:30:42,880
It's just good to know where it's going to go.

420
00:30:42,880 --> 00:30:49,160
Yeah, very much so, which I think, sorry, I cut you off there, but that leads us neatly

421
00:30:49,160 --> 00:31:01,280
into the next story and a model which is very rooted in transparency and responsible usage

422
00:31:01,280 --> 00:31:02,960
and respect for its community.

423
00:31:02,960 --> 00:31:08,080
I think if you read the rules around this next story, you'll see that the Photoshop

424
00:31:08,080 --> 00:31:13,600
are doing a lot to, sorry, I've completely balls that up, not Photoshop, Adobe have done

425
00:31:13,600 --> 00:31:15,200
a lot to respect creators.

426
00:31:15,200 --> 00:31:17,840
So with that in mind, Paul, what were you going to say?

427
00:31:17,840 --> 00:31:18,840
Because I cut you off.

428
00:31:18,840 --> 00:31:20,280
That's a bit of a ramble.

429
00:31:20,280 --> 00:31:23,560
Welcome listeners to my mind.

430
00:31:23,560 --> 00:31:29,280
We promise as much as we wish we hadn't done this to ourselves, we're not going to cut

431
00:31:29,280 --> 00:31:35,320
any of the mistakes because as we've talked about, the human authenticity of this content

432
00:31:35,320 --> 00:31:41,040
is absolutely critical so that you know that we're not bots.

433
00:31:41,040 --> 00:31:44,360
Although would bots deliberately make those mistakes and then leave them in?

434
00:31:44,360 --> 00:31:47,160
Oh dear listener, we'll leave you to chew on that.

435
00:31:47,160 --> 00:31:52,400
But yeah, let's make a segue and then take this into our next story, which is Photoshop's

436
00:31:52,400 --> 00:31:56,480
generative AI beta, which has been launched.

437
00:31:56,480 --> 00:32:00,320
We do have an Adobe account here at Biostratus where I was able to have a play with it.

438
00:32:00,320 --> 00:32:04,360
It was surprisingly easy to access, although you had to update Photoshop and then you had

439
00:32:04,360 --> 00:32:08,440
to opt into the beta and then you had to update the beta part of Photoshop and then you had

440
00:32:08,440 --> 00:32:10,240
to update Photoshop again.

441
00:32:10,240 --> 00:32:14,840
It wasn't an easy process to get hold of it, but we were able to and we've had a play with

442
00:32:14,840 --> 00:32:15,840
it.

443
00:32:15,840 --> 00:32:16,840
Now, what is this?

444
00:32:16,840 --> 00:32:20,960
So Photoshop has introduced this new feature called generative fill.

445
00:32:20,960 --> 00:32:28,840
It basically allows users to access the Firefly tool that they've had in beta as like a separate

446
00:32:28,840 --> 00:32:33,700
standalone and some of the power of that directly in Photoshop when they're working in their

447
00:32:33,700 --> 00:32:36,640
normal image workflow.

448
00:32:36,640 --> 00:32:42,800
Using generative fill, people can use natural language to guide Photoshop in creating aspects

449
00:32:42,800 --> 00:32:43,800
of an image.

450
00:32:43,800 --> 00:32:44,800
So what can they create?

451
00:32:44,800 --> 00:32:51,440
They can add elements, they can replace parts of an image, they can expand past the edges

452
00:32:51,440 --> 00:32:55,520
of an image or, you know, basically at the top or just make an image bigger and it sort

453
00:32:55,520 --> 00:32:59,520
of auto fills in a very contextually aware way.

454
00:32:59,520 --> 00:33:02,560
There's loads of good stuff that you can do.

455
00:33:02,560 --> 00:33:04,200
I've been playing with it this week.

456
00:33:04,200 --> 00:33:09,200
I've had a couple of my very good friends having a play with it as well.

457
00:33:09,200 --> 00:33:14,200
Emile Lampreck, who we've had on the podcast before and also Andrew Muir, props to you

458
00:33:14,200 --> 00:33:19,000
guys for having a proper WhatsApp bash of us just sharing the different images and what

459
00:33:19,000 --> 00:33:21,240
worked well and what didn't.

460
00:33:21,240 --> 00:33:25,640
And at first, I've got to admit, I wasn't super impressed because I tried some things

461
00:33:25,640 --> 00:33:27,680
and it just did a very poor job of it.

462
00:33:27,680 --> 00:33:34,960
So there's some skill to figuring out how to use the prompt, like unlike, say, Mid Journey

463
00:33:34,960 --> 00:33:42,420
or Chat GPT, if you're overly instructional, in my experience at least, it pushes the instruction

464
00:33:42,420 --> 00:33:44,740
into the output.

465
00:33:44,740 --> 00:33:51,040
So I made a picture of a horse and then I said, add a tattoo of a lion, but then it

466
00:33:51,040 --> 00:33:58,440
put a person adding the tattoo onto the lion and I just wanted the tattoo.

467
00:33:58,440 --> 00:34:04,160
So when I said add, it included a person, a tattoo artist, not the tattoo.

468
00:34:04,160 --> 00:34:10,280
Anyway, did you manage to fix that and maybe just put lion tattoo?

469
00:34:10,280 --> 00:34:11,280
And that's exactly what I did.

470
00:34:11,280 --> 00:34:15,880
And that's exactly what then worked pretty poorly, honestly, but that was one of my early

471
00:34:15,880 --> 00:34:17,160
attempts.

472
00:34:17,160 --> 00:34:18,160
Some of the things that we found...

473
00:34:18,160 --> 00:34:20,480
It's difficult to tattoo hair.

474
00:34:20,480 --> 00:34:22,160
Do you think this was the...

475
00:34:22,160 --> 00:34:25,000
Hey, I just thought this is an absurd request.

476
00:34:25,000 --> 00:34:26,000
I refuse.

477
00:34:26,000 --> 00:34:27,600
It's a stupid idea.

478
00:34:27,600 --> 00:34:28,600
Yeah.

479
00:34:28,600 --> 00:34:31,680
I mean, it was a stupid idea and it didn't work very well.

480
00:34:31,680 --> 00:34:37,160
With more playing between Emil, Andrew and myself, we got it to do some pretty cool stuff.

481
00:34:37,160 --> 00:34:45,360
So Andrew spent some time extracting Emil and I from other images with the Context-Aware

482
00:34:45,360 --> 00:34:47,320
Select tool, which is awesome.

483
00:34:47,320 --> 00:34:51,760
It's kind of like, you know, we looked at that segmentation model from Meta that was

484
00:34:51,760 --> 00:34:53,160
really powerful for just...

485
00:34:53,160 --> 00:34:57,560
You click on a thing and it knows in a smart way where the boundaries are.

486
00:34:57,560 --> 00:35:03,440
The sort of magnetic lasso is going to die a death because you can just click on a person

487
00:35:03,440 --> 00:35:08,200
and it very intelligently selects them and then you can remove the background really

488
00:35:08,200 --> 00:35:11,360
quickly and then chuck other backgrounds in.

489
00:35:11,360 --> 00:35:18,160
So Andrew had poor Emil was being attacked by the Loch Ness Monster, a bear.

490
00:35:18,160 --> 00:35:24,040
At one point we were trying to see if we could get Ridley Scott's alien in there.

491
00:35:24,040 --> 00:35:29,400
Basically both Emil and I and Andrew's images were under siege by all these different characters.

492
00:35:29,400 --> 00:35:34,000
And in many cases, we're actually able to get some really quite good effects.

493
00:35:34,000 --> 00:35:37,440
Obviously didn't necessarily look real.

494
00:35:37,440 --> 00:35:42,240
But the bear, for example, was completely in the context of the background, even though

495
00:35:42,240 --> 00:35:47,760
the area where the bear was inserted wasn't in the original image.

496
00:35:47,760 --> 00:35:51,880
Photoshop had to create the background and then put the bear in it.

497
00:35:51,880 --> 00:35:55,800
And it got it sort of on the ground, if you know what I mean, like stood in the right

498
00:35:55,800 --> 00:35:56,800
place.

499
00:35:56,800 --> 00:35:57,800
It was pretty cool.

500
00:35:57,800 --> 00:36:04,000
In fact, the Context-Aware Expansion, so taking like a four by three image and expanding it

501
00:36:04,000 --> 00:36:09,720
to 16.9, like you would do, say for a blog post, Eder, or maybe when you're fiddling

502
00:36:09,720 --> 00:36:14,080
with images for different formats on social or different ad formats.

503
00:36:14,080 --> 00:36:16,360
That was also surprisingly good.

504
00:36:16,360 --> 00:36:22,680
Emil ran an incredibly complex requirement where the centralized image to expand out

505
00:36:22,680 --> 00:36:25,480
was going to be really difficult and it did a really good job.

506
00:36:25,480 --> 00:36:31,200
So I think if you're a marketer, you should know about this because it's going to bring

507
00:36:31,200 --> 00:36:36,240
the power to you that usually would have required some significant design expertise.

508
00:36:36,240 --> 00:36:41,800
I think if you're a designer, you could probably use this to speed up your workflow significantly

509
00:36:41,800 --> 00:36:46,160
and it's absolutely worth having a play with.

510
00:36:46,160 --> 00:36:52,560
This came hot on the heels of a research paper about a tool called Dragon, Dragon, Dragon,

511
00:36:52,560 --> 00:36:55,840
I think it's Dragon, that came out of the Max Planck Institute.

512
00:36:55,840 --> 00:36:57,240
It's not publicly available yet.

513
00:36:57,240 --> 00:36:59,120
It was a paper.

514
00:36:59,120 --> 00:37:03,840
You've probably seen it on social media if you've clicked on anything AI related in the

515
00:37:03,840 --> 00:37:07,600
last few months because I've seen it in many different places.

516
00:37:07,600 --> 00:37:09,280
The demo video is pretty impressive.

517
00:37:09,280 --> 00:37:16,280
It shows an image editor a bit like Photoshop, but where you can drag around and it will

518
00:37:16,280 --> 00:37:18,360
contextually aware edit the image.

519
00:37:18,360 --> 00:37:24,280
So an example was you could close and open the eyes of a cat and it looked at every stage

520
00:37:24,280 --> 00:37:25,880
basically real.

521
00:37:25,880 --> 00:37:30,760
You could open a lion's mouth and turn their head and it was contextually aware enough

522
00:37:30,760 --> 00:37:35,000
to again make it look like that's how the original image had been captured.

523
00:37:35,000 --> 00:37:40,080
There was a resizing of a car where you could change the size of the wheels or the trunk

524
00:37:40,080 --> 00:37:45,120
and actually the image would adjust, like the wheel arches would move to accommodate

525
00:37:45,120 --> 00:37:46,600
the new larger wheels.

526
00:37:46,600 --> 00:37:47,920
Like pretty cool.

527
00:37:47,920 --> 00:37:52,320
You could adjust a person's clothing like the lengths of their sleeves or a skirt or

528
00:37:52,320 --> 00:37:57,040
trousers and things like that and it would again contextually aware and it was able to

529
00:37:57,040 --> 00:38:00,800
do that with really surprising fidelity and make it look real.

530
00:38:00,800 --> 00:38:08,760
So loads of really interesting image adjustment news this week, which of course is absolutely

531
00:38:08,760 --> 00:38:13,080
critical for marketers in a huge amount of areas in which they work.

532
00:38:13,080 --> 00:38:16,640
Yeah, and it's amazing to see the development in this area.

533
00:38:16,640 --> 00:38:27,080
I was showing off the Canva Magic Edit, Magic Eraser feature, which is basically in-painting

534
00:38:27,080 --> 00:38:31,320
much like the tools that you've been describing with Photoshop.

535
00:38:31,320 --> 00:38:36,840
Although the Photoshop ones obviously have a wider suite attached to them, but yeah,

536
00:38:36,840 --> 00:38:37,840
it's cool.

537
00:38:37,840 --> 00:38:43,360
These are readily available to anybody to use and you really don't need to be a particularly

538
00:38:43,360 --> 00:38:52,160
skilled designer, artist, illustrator to be able to get something that looks pretty decent.

539
00:38:52,160 --> 00:38:54,960
The democratization of design, eh?

540
00:38:54,960 --> 00:38:56,120
Absolutely.

541
00:38:56,120 --> 00:39:02,880
And I still think the demo videos always show you the very best possible examples.

542
00:39:02,880 --> 00:39:10,560
I had an image that I used in one of our blogs last week of a little bit of satire.

543
00:39:10,560 --> 00:39:16,480
It was robots sitting in the Houses of Parliament writing up the AI Act, but doing it with paper

544
00:39:16,480 --> 00:39:24,560
and pen, which I just felt was apt in terms of how government institutions might use AI

545
00:39:24,560 --> 00:39:29,160
to increase efficiency in some areas and then still make everything work with pen and paper,

546
00:39:29,160 --> 00:39:31,000
but I digress.

547
00:39:31,000 --> 00:39:38,440
When I tried to swap that pen for a feather quill, it just looked like I'd use an image

548
00:39:38,440 --> 00:39:42,240
stamp like it was not contextually aware at all.

549
00:39:42,240 --> 00:39:44,640
I tried to put a hat on one of the robots.

550
00:39:44,640 --> 00:39:46,360
Awful, didn't work at all.

551
00:39:46,360 --> 00:39:51,680
But if you watch the demo video, it's contextually aware enough, at least in certain cases, to

552
00:39:51,680 --> 00:39:57,600
figure out what the light source is and show light bouncing and reflections in rivers and

553
00:39:57,600 --> 00:40:03,560
stuff like that with a decent level of accuracy that again, I think you can get, but you've

554
00:40:03,560 --> 00:40:09,080
really got to get the prompt right, make the selection be just in the right place, try

555
00:40:09,080 --> 00:40:14,080
different prompts, and I think some images are just going to work better than others.

556
00:40:14,080 --> 00:40:16,800
So yeah, great power.

557
00:40:16,800 --> 00:40:21,920
I think there's cool stuff that we can do, still frustrating enough at the moment to

558
00:40:21,920 --> 00:40:25,960
maybe drive you a bit mad if you've got a particular vision in your head of how it should

559
00:40:25,960 --> 00:40:27,200
look.

560
00:40:27,200 --> 00:40:29,600
I still think it's hard to get that.

561
00:40:29,600 --> 00:40:36,720
That reminds me that I was playing with Firefly this week with the text to effect or the text

562
00:40:36,720 --> 00:40:37,720
effect tool.

563
00:40:37,720 --> 00:40:43,040
I know that you'd said that was somewhat frustrating when we had that conversation.

564
00:40:43,040 --> 00:40:46,720
It can't quite do certain textures.

565
00:40:46,720 --> 00:40:55,280
This week I tried to get a text effect with Donna Kebab and didn't work very well.

566
00:40:55,280 --> 00:41:00,280
If anything, the reason my critique of it was that it was putting in a Sheik Kebab,

567
00:41:00,280 --> 00:41:02,480
which that's so wrong.

568
00:41:02,480 --> 00:41:06,400
It's a million miles away, idiot AI.

569
00:41:06,400 --> 00:41:11,680
I mean, there's a support ticket there.

570
00:41:11,680 --> 00:41:12,680
Get on there.

571
00:41:12,680 --> 00:41:13,680
Right.

572
00:41:13,680 --> 00:41:18,560
Let's move on to our last main story of the week, which is on a similar vein actually,

573
00:41:18,560 --> 00:41:22,320
Google adds generative AI to ads and image assets.

574
00:41:22,320 --> 00:41:25,760
Myan, do you want to tell us about this one briefly?

575
00:41:25,760 --> 00:41:26,760
Yeah.

576
00:41:26,760 --> 00:41:32,080
So a very practical use case, kind of obvious that Google would launch this.

577
00:41:32,080 --> 00:41:36,400
It's going to help people create ads in the AdWords platform.

578
00:41:36,400 --> 00:41:42,640
So they're bringing generative AI to the AdWords platform to simply add, basically you simply

579
00:41:42,640 --> 00:41:49,720
add your preferred landing page from your website and then using the text and the context

580
00:41:49,720 --> 00:41:54,280
from the landing page, Google will then create your ads.

581
00:41:54,280 --> 00:41:59,640
So it will generate relevant keywords, headlines, descriptions, images, and basically pull in

582
00:41:59,640 --> 00:42:05,760
any other assets and using their huge amounts of knowledge about what is effective in terms

583
00:42:05,760 --> 00:42:13,320
of headlines and copywriting and what have you, they suggest that these will be rather

584
00:42:13,320 --> 00:42:16,880
effective combinations of ad creatives.

585
00:42:16,880 --> 00:42:21,760
Maybe you can review and edit these if you wanted to say something different, if you're

586
00:42:21,760 --> 00:42:23,680
not happy with it.

587
00:42:23,680 --> 00:42:29,000
But yeah, you can now get AI to put together your campaign.

588
00:42:29,000 --> 00:42:32,440
Now of course, you've still got to have a landing page that you create.

589
00:42:32,440 --> 00:42:36,200
You've got to have a page that gives good context and has a compelling offer in order

590
00:42:36,200 --> 00:42:37,360
to get the conversions.

591
00:42:37,360 --> 00:42:43,480
But then once you transfer that over to AdWords, you'll be good to go.

592
00:42:43,480 --> 00:42:44,480
Yeah.

593
00:42:44,480 --> 00:42:45,480
Interesting.

594
00:42:45,480 --> 00:42:51,680
I think we've talked a little bit about programmatic and advertising recently.

595
00:42:51,680 --> 00:42:58,480
We've talked a bit about AI enabled ad management and sort of auto creation of ad text and stuff

596
00:42:58,480 --> 00:43:00,880
like that.

597
00:43:00,880 --> 00:43:03,520
Keyword selection, automatic bidding and all of that stuff.

598
00:43:03,520 --> 00:43:08,280
So this obviously fits in line with that conversation and will certainly speed up workflows.

599
00:43:08,280 --> 00:43:15,400
When I saw this, I'm a member of a number of direct response marketing Facebook groups

600
00:43:15,400 --> 00:43:18,840
which is kind of very different to the B2B world that I usually play in.

601
00:43:18,840 --> 00:43:22,800
It's very much consumer oriented, a lot of it's informational products.

602
00:43:22,800 --> 00:43:27,840
But I find it fascinating to see some of the conversations because it's a lot of it's

603
00:43:27,840 --> 00:43:37,640
proper old school, emotive, emotion based, persuasive copywriting.

604
00:43:37,640 --> 00:43:41,960
The type of which has fallen out of a lot of marketing over the years, certainly in

605
00:43:41,960 --> 00:43:47,680
B2B or technical industries like the life sciences where despite our best efforts, everything

606
00:43:47,680 --> 00:43:52,280
ends up very feature driven and first order benefits.

607
00:43:52,280 --> 00:43:56,640
If you're lucky, never second order benefits or what it feels like to finally get a nature

608
00:43:56,640 --> 00:44:02,080
paper or to be able to unlock a new research application that wasn't possible before.

609
00:44:02,080 --> 00:44:08,160
When I read those, I really get a sense for the world class copywriters in the world who

610
00:44:08,160 --> 00:44:17,480
understand human psychology, understand what motivates people, understand true persuasion,

611
00:44:17,480 --> 00:44:24,440
write copy and structure landing pages and add copy and stuff in with real skill.

612
00:44:24,440 --> 00:44:29,660
And I'm not saying that AI couldn't learn to do that if trained on enough examples,

613
00:44:29,660 --> 00:44:34,540
but my fear with this would be it's a great way to get ads up quickly.

614
00:44:34,540 --> 00:44:37,880
Is it a great way to get the best convert in ads?

615
00:44:37,880 --> 00:44:42,080
Is it the best way to get the best results long term?

616
00:44:42,080 --> 00:44:43,320
I don't know.

617
00:44:43,320 --> 00:44:47,480
I think that the world class copywriters of the world would say no.

618
00:44:47,480 --> 00:44:51,440
And certainly a bunch of stuff that I've been able to get out of generative tools like chat

619
00:44:51,440 --> 00:44:55,000
GPT is not bad.

620
00:44:55,000 --> 00:44:59,220
And if you tell it to think in a certain way and use a particular marketing framework like

621
00:44:59,220 --> 00:45:04,800
Ada or what have you, it can do a decent job, but nothing compared to some of these really

622
00:45:04,800 --> 00:45:09,040
gifted direct response marketing copywriters.

623
00:45:09,040 --> 00:45:12,560
So yeah, I think that would be the one caveat I would put on it.

624
00:45:12,560 --> 00:45:14,600
Yeah, a reasonable concern.

625
00:45:14,600 --> 00:45:20,640
And well, I guess we'll have to wait and see.

626
00:45:20,640 --> 00:45:28,240
What ultimately Google and Meta as well are also getting into the generative AI assistant

627
00:45:28,240 --> 00:45:31,680
for ad creation.

628
00:45:31,680 --> 00:45:43,160
I guess the good thing or the thing that they have versus a copywriter is a really good

629
00:45:43,160 --> 00:45:51,440
copywriter can spend time crafting say one, two, three, maybe a handful of variants.

630
00:45:51,440 --> 00:46:00,240
These tools, if you kind of hand them over, presumably we'll be able to multivariant test

631
00:46:00,240 --> 00:46:06,160
different combinations, different phrases at scale automatically and find combinations

632
00:46:06,160 --> 00:46:14,200
that quite frankly, the human just wouldn't have, you know, take them days, weeks, months

633
00:46:14,200 --> 00:46:15,200
to do.

634
00:46:15,200 --> 00:46:16,200
I'd agree.

635
00:46:16,200 --> 00:46:20,340
I think if you've got a large enough budget and a large enough audience in terms of search

636
00:46:20,340 --> 00:46:23,400
volume around those keywords, it's really funny because in the live science, there's

637
00:46:23,400 --> 00:46:26,580
a lot of the keywords that we deal with a quite low search volume and we have to apply

638
00:46:26,580 --> 00:46:32,120
very specific strategies and tactics around that, which this wouldn't work for, right?

639
00:46:32,120 --> 00:46:33,120
Potentially.

640
00:46:33,120 --> 00:46:39,320
I also think to your point, if you're able to produce those, there's nothing to stop

641
00:46:39,320 --> 00:46:43,080
an expert reviewing them and then choosing which they think is going to be the best based

642
00:46:43,080 --> 00:46:44,080
on their experience.

643
00:46:44,080 --> 00:46:49,520
So again, experts human in the loop augmented to do this at scale and speed is probably

644
00:46:49,520 --> 00:46:54,200
going to end up being the way forward at least to begin with.

645
00:46:54,200 --> 00:46:56,800
So from that perspective, I think it makes a lot of sense.

646
00:46:56,800 --> 00:47:04,360
And ultimately, if the main play here outside of, I think your very good point about being

647
00:47:04,360 --> 00:47:08,280
able to split test and find out what works the best.

648
00:47:08,280 --> 00:47:12,340
But if outside of that, the main players speed and efficiency, well, what does that do?

649
00:47:12,340 --> 00:47:18,640
It means that there's ever more campaigns being running, loads more ads, loads more

650
00:47:18,640 --> 00:47:20,640
spend all the things that Google wants.

651
00:47:20,640 --> 00:47:24,400
But as an advertiser is not necessarily what I want because now there's more competition

652
00:47:24,400 --> 00:47:26,800
and click costs go up.

653
00:47:26,800 --> 00:47:31,280
I'm under even more pressure now to get the best cost per click, the best click through

654
00:47:31,280 --> 00:47:33,140
rates, the best conversion rate.

655
00:47:33,140 --> 00:47:37,840
And now we're back to having a really strong value proposition, effective messaging that's

656
00:47:37,840 --> 00:47:42,920
laser targeted on the interests of your customers, which is usually best done through primary

657
00:47:42,920 --> 00:47:46,680
research, which of course, AIs can't do yet.

658
00:47:46,680 --> 00:47:50,600
And then probably that's where the benefit of world class copywriting to really make

659
00:47:50,600 --> 00:47:55,040
sure that your landing pages and ads and things stand out.

660
00:47:55,040 --> 00:47:58,800
So it could end up being like quite a weird cycle in some ways.

661
00:47:58,800 --> 00:48:02,360
Yeah, no, a hundred percent.

662
00:48:02,360 --> 00:48:03,360
Right.

663
00:48:03,360 --> 00:48:04,360
Talking of weird cycles.

664
00:48:04,360 --> 00:48:07,840
No, there is no weird cycle to link onto.

665
00:48:07,840 --> 00:48:15,240
Talking of all good things, marketing, automation and cool generative AI and other tools.

666
00:48:15,240 --> 00:48:17,640
We have been playing, having with Zapier, Martin.

667
00:48:17,640 --> 00:48:21,600
So that's going to be our tool of the week this week and some of the cool stuff we've

668
00:48:21,600 --> 00:48:22,840
been doing with it.

669
00:48:22,840 --> 00:48:25,680
I know we've had both some experiences this week.

670
00:48:25,680 --> 00:48:28,880
Why don't you tell us what you've been up to, Martin?

671
00:48:28,880 --> 00:48:38,840
I finally connected my OpenAI API, a key with Zapier and started just playing around with

672
00:48:38,840 --> 00:48:40,960
a few multi-step zaps.

673
00:48:40,960 --> 00:48:45,640
So if people aren't playing with Zapier already to automate standard marketing tasks in their

674
00:48:45,640 --> 00:48:48,840
day-to-day lives, then they absolutely should get on it.

675
00:48:48,840 --> 00:48:56,840
But the first thing I noticed when I looked at what you can do with the OpenAI actions

676
00:48:56,840 --> 00:49:02,760
and the triggers, I noticed that it has access to the Whisper API, which is the speech to

677
00:49:02,760 --> 00:49:05,480
text transcription tool.

678
00:49:05,480 --> 00:49:08,400
Now that's something that I've not been able to play with.

679
00:49:08,400 --> 00:49:09,400
I'm not a coder.

680
00:49:09,400 --> 00:49:12,560
I don't know how to actually build on the API itself.

681
00:49:12,560 --> 00:49:16,280
This gave me a no-code solution for doing that.

682
00:49:16,280 --> 00:49:25,440
So what I did was I made a zap that starts off by monitoring a Google Drive folder.

683
00:49:25,440 --> 00:49:33,280
Every time a new file goes into there, it will upload that file to the Whisper API.

684
00:49:33,280 --> 00:49:40,640
So I will save a voice note from my phone into my Google Drive folder.

685
00:49:40,640 --> 00:49:43,040
It then puts that into Whisper and takes the transcript.

686
00:49:43,040 --> 00:49:52,720
I've then created a ChatGPT prompt, which inserts the transcript into ChatGPT, grabs

687
00:49:52,720 --> 00:50:00,840
the response from ChatGPT, which is a blog post based around the audio recording that

688
00:50:00,840 --> 00:50:03,440
I transcribed.

689
00:50:03,440 --> 00:50:08,580
That then spits that out into a Google Doc and saves it in a separate folder.

690
00:50:08,580 --> 00:50:15,720
So by recording a very short voice note, uploading that to Google Drive, I get a blog post created.

691
00:50:15,720 --> 00:50:21,920
I did this in a live demo with a group this week, a two-minute, five-question interview

692
00:50:21,920 --> 00:50:26,240
with a cybersecurity specialist.

693
00:50:26,240 --> 00:50:27,240
We went back and forth.

694
00:50:27,240 --> 00:50:29,140
I asked them some very simple questions.

695
00:50:29,140 --> 00:50:37,540
At the end of it, we had a very neat 500-word blog post created, which with very light touch

696
00:50:37,540 --> 00:50:41,840
editing and a little bit of expansion on some of the points, we could have easily got to

697
00:50:41,840 --> 00:50:48,400
750, 1,000 words long that covered some of the big issues in the industry.

698
00:50:48,400 --> 00:50:53,440
All in all, I think we could have had a blog post created and written with two minutes

699
00:50:53,440 --> 00:50:59,660
worth of recording, a minute's worth of uploading to Google Drive, and then maybe 20 minutes

700
00:50:59,660 --> 00:51:02,840
of relatively light editing.

701
00:51:02,840 --> 00:51:03,840
Yeah.

702
00:51:03,840 --> 00:51:07,760
I love that use case, Martin.

703
00:51:07,760 --> 00:51:14,360
I think what I love about Zapier is providing us with the tools to be able to come up with

704
00:51:14,360 --> 00:51:20,200
cool applications for ourselves without having to code.

705
00:51:20,200 --> 00:51:25,320
Because what you did there is absolute genius, and then Zapier makes that easy by connecting

706
00:51:25,320 --> 00:51:27,980
all the different tools together.

707
00:51:27,980 --> 00:51:35,460
This week, I was on a webinar called How Zapier's Go-to-Market Leaders Are Using AI.

708
00:51:35,460 --> 00:51:41,040
There was just a real bunch of interesting use cases in that webinar.

709
00:51:41,040 --> 00:51:43,680
I recommend Googling it and get me on to it.

710
00:51:43,680 --> 00:51:47,800
Maybe we can even include the link in the show notes, I'm not sure.

711
00:51:47,800 --> 00:51:52,760
I'm watching it because I think it will really open your mind to what's possible.

712
00:51:52,760 --> 00:51:58,880
A huge amount of what they're doing is very Slack oriented.

713
00:51:58,880 --> 00:52:03,520
They'll have it so that, for example, for their sales team, they will be able to ask

714
00:52:03,520 --> 00:52:10,520
for information on a particular customer from a Slack bot that then through a mixture of

715
00:52:10,520 --> 00:52:15,860
tools in the backend and different data sources it's connect to understands natural language

716
00:52:15,860 --> 00:52:18,360
questions about the customer.

717
00:52:18,360 --> 00:52:23,840
Surfaces information about their business, if it's B2B, also reports information back

718
00:52:23,840 --> 00:52:28,040
from the CRM on different parts of, in this case, the software that they've used recently

719
00:52:28,040 --> 00:52:32,080
and all those types of things, and actually just gives them that little customer summary

720
00:52:32,080 --> 00:52:34,600
ready to go into the sales meeting.

721
00:52:34,600 --> 00:52:41,160
But it's all done sort of automagically, but through natural language bot interface.

722
00:52:41,160 --> 00:52:44,920
It's almost like having a sales assistant that you're like, right, I've got to speak

723
00:52:44,920 --> 00:52:47,280
to Bob from Acme Co.

724
00:52:47,280 --> 00:52:51,280
Tell me about Bob and then the chat bot just explains it all to you, which I just thought

725
00:52:51,280 --> 00:52:52,760
was really interesting.

726
00:52:52,760 --> 00:52:56,680
I saw a similar one that someone had done using HubSpot.

727
00:52:56,680 --> 00:52:57,680
Oh really?

728
00:52:57,680 --> 00:52:58,680
Yeah.

729
00:52:58,680 --> 00:53:03,840
So because you can connect HubSpot with Zapier, this was on LinkedIn, someone shared this,

730
00:53:03,840 --> 00:53:08,400
and it was a sales outreach email and they said, really neat because the amount of data

731
00:53:08,400 --> 00:53:13,680
that you can extract from HubSpot in Zapier, all of the fields are basically available.

732
00:53:13,680 --> 00:53:20,640
So every time a new lead was generated, you can have that go into the contact, extract

733
00:53:20,640 --> 00:53:21,920
all of the relevant information.

734
00:53:21,920 --> 00:53:27,920
And they pulled through in this prompt, they had, I think it was about 20 fields.

735
00:53:27,920 --> 00:53:33,240
So they had the basics of first name, last name, email address, job title, et cetera.

736
00:53:33,240 --> 00:53:40,960
But because HubSpot has HubSpot Insights, which is automatically populated data about the

737
00:53:40,960 --> 00:53:45,960
company of someone who's been on your website and filled out a form, it was able to extract

738
00:53:45,960 --> 00:53:47,960
all of that information as well.

739
00:53:47,960 --> 00:53:52,720
So this was what the company is, what the kind of annual revenue is, the kind of estimated

740
00:53:52,720 --> 00:53:57,820
annual revenue, number of employees, the city that it's based in, kind of one paragraph

741
00:53:57,820 --> 00:54:02,520
description of the industry that it's in, its website domain, all of that kind of stuff

742
00:54:02,520 --> 00:54:04,620
goes into the prompt.

743
00:54:04,620 --> 00:54:08,400
And then it said, oh yeah, what page they submitted the form on.

744
00:54:08,400 --> 00:54:13,400
So it was contextually aware about what the context of their form submission was.

745
00:54:13,400 --> 00:54:16,840
So maybe it was a webinar or white paper or something.

746
00:54:16,840 --> 00:54:23,160
And then got it to wrote a sales email, pushed that through into a follow-up email, pushed

747
00:54:23,160 --> 00:54:26,240
that into Gmail and send it.

748
00:54:26,240 --> 00:54:27,240
That's awesome.

749
00:54:27,240 --> 00:54:30,240
It was, I watched it and I thought that is super impressive.

750
00:54:30,240 --> 00:54:35,560
Like it's obvious when you see it, but it worked really neatly.

751
00:54:35,560 --> 00:54:36,800
That makes sense.

752
00:54:36,800 --> 00:54:40,080
I wonder if I'm not a big Zapier user yet.

753
00:54:40,080 --> 00:54:42,000
I've had an account for many years.

754
00:54:42,000 --> 00:54:45,400
Obviously this has prompted me to play with it quite a bit more after we were talking

755
00:54:45,400 --> 00:54:50,880
about it through the WeMind, but can you link LinkedIn sales navigator?

756
00:54:50,880 --> 00:54:55,360
Basically link everything to Zapier pretty much, kind of, because then you could pull

757
00:54:55,360 --> 00:54:59,800
things on an individual into that, not just their company, right?

758
00:54:59,800 --> 00:55:01,520
Yeah, good question.

759
00:55:01,520 --> 00:55:08,240
I don't know why not.

760
00:55:08,240 --> 00:55:12,800
Be worth having a quick Google.

761
00:55:12,800 --> 00:55:14,280
It looks like there is an integration.

762
00:55:14,280 --> 00:55:18,160
Oh no, LinkedIn sales navigator has not yet built an integration on Zapier.

763
00:55:18,160 --> 00:55:19,160
Boo right.

764
00:55:19,160 --> 00:55:20,160
We'd love that please.

765
00:55:20,160 --> 00:55:21,160
That would be pretty awesome.

766
00:55:21,160 --> 00:55:26,600
Another thing that was in the webinar that was pretty cool was they'd noticed that GPT

767
00:55:26,600 --> 00:55:33,400
based tools are really good at summarizing sales calls, but also classifying them.

768
00:55:33,400 --> 00:55:35,720
What were the main features?

769
00:55:35,720 --> 00:55:37,480
Because obviously Zapier's team sells software.

770
00:55:37,480 --> 00:55:40,120
What were the main features that the lead was interested in?

771
00:55:40,120 --> 00:55:42,560
What were the objections that they came up with?

772
00:55:42,560 --> 00:55:49,640
And then building a classification mechanism for GPT to draw upon to then auto classify

773
00:55:49,640 --> 00:55:50,640
those.

774
00:55:50,640 --> 00:55:54,340
Which again, if you're doing all of these things at the scale that some of these enterprise

775
00:55:54,340 --> 00:56:00,480
companies would be, there's probably loads of insights you could pull from having that

776
00:56:00,480 --> 00:56:06,040
quick summary of that type of information, which I thought was really interesting.

777
00:56:06,040 --> 00:56:12,200
So I think if you're not using Zapier to explore some of this stuff yet, you really should

778
00:56:12,200 --> 00:56:13,200
be.

779
00:56:13,200 --> 00:56:14,840
That's certainly how I feel.

780
00:56:14,840 --> 00:56:17,360
And Martin, your stories are even more inspiring.

781
00:56:17,360 --> 00:56:22,840
I guess in some cases, you're going to need to have API access to some of these tools

782
00:56:22,840 --> 00:56:24,000
to enable this, right?

783
00:56:24,000 --> 00:56:30,640
The OpenAI example, the chat GPT example you gave, that wasn't chat GPT.

784
00:56:30,640 --> 00:56:35,000
That was access to GPT-4 API, one assumes, to be able to do that.

785
00:56:35,000 --> 00:56:36,000
Yeah.

786
00:56:36,000 --> 00:56:38,720
So actually I don't have GPT-4 API access.

787
00:56:38,720 --> 00:56:45,760
That just came from going into platform.openai.com, making sure I could generate an API key.

788
00:56:45,760 --> 00:56:54,400
And I created a chat GPT API key, which is limited to chat GPT 3.5 or 3.5 turbo.

789
00:56:54,400 --> 00:56:55,400
Right.

790
00:56:55,400 --> 00:56:56,400
Okay.

791
00:56:56,400 --> 00:56:57,880
But yeah, but you need API access.

792
00:56:57,880 --> 00:56:58,880
Yeah, yeah.

793
00:56:58,880 --> 00:57:01,120
The standard account won't do it.

794
00:57:01,120 --> 00:57:03,480
You do need to create a developer account.

795
00:57:03,480 --> 00:57:04,480
Cool.

796
00:57:04,480 --> 00:57:08,640
The other thing that's worth looking at is Zapier have now got this new interfaces tool,

797
00:57:08,640 --> 00:57:15,160
which is in beta, which promises to allow you to build forms, web pages, and basic apps.

798
00:57:15,160 --> 00:57:24,760
No coding required, but where you can connect to GPT-based models like GPT 3.5.

799
00:57:24,760 --> 00:57:33,040
So in essence, it has an element of chat bot builder about it, but this is the underlying

800
00:57:33,040 --> 00:57:37,080
tool that they're using for a lot of the use cases we've been describing, because that's

801
00:57:37,080 --> 00:57:43,480
how they end up basically then this lives in Slack, but its ability to draw on all your

802
00:57:43,480 --> 00:57:48,760
other data sources through all your other Zapier connections is where the power is.

803
00:57:48,760 --> 00:57:52,200
So a lot of the ones we talked about were actually internal Zapier tools, right?

804
00:57:52,200 --> 00:57:56,800
Being used internally by their teams, but one assumes you can actually create customer

805
00:57:56,800 --> 00:58:04,760
portals or bots on your website that could perform similar actions, but low stroke, no

806
00:58:04,760 --> 00:58:08,840
code, because you're just doing it through Zapier instead of having to figure out yourself

807
00:58:08,840 --> 00:58:13,600
how to connect different tools and move data from one tool to the next tool and all that

808
00:58:13,600 --> 00:58:14,600
other stuff.

809
00:58:14,600 --> 00:58:20,880
So in some ways, having access to generative AI is like a massive accelerator for all those

810
00:58:20,880 --> 00:58:24,800
Zapier data connections that you set up over the years, but don't really bother to build

811
00:58:24,800 --> 00:58:28,760
many Zapier apps for anymore, if that's just me.

812
00:58:28,760 --> 00:58:29,760
Cool.

813
00:58:29,760 --> 00:58:33,880
Well, hopefully that was a useful episode for everyone.

814
00:58:33,880 --> 00:58:37,600
If you enjoyed, please subscribe, share it with your marketing friends.

815
00:58:37,600 --> 00:58:42,160
If there's things you'd love us to cover on the podcast that we don't, hit us up on the

816
00:58:42,160 --> 00:58:45,540
Twitters, connect with us and chat with us on LinkedIn.

817
00:58:45,540 --> 00:58:48,640
If there's things that we talk about that you're like, nah, I actually think that bit

818
00:58:48,640 --> 00:58:49,640
of the podcast boring.

819
00:58:49,640 --> 00:58:51,380
Well, we want to hear that as well, right?

820
00:58:51,380 --> 00:58:54,040
Because all feedback is very welcome.

821
00:58:54,040 --> 00:58:57,320
If you want to come on as an interviewee, because you've got something interesting to

822
00:58:57,320 --> 00:59:02,380
say, maybe you've been piloting some cool AI driven applications of your own, maybe

823
00:59:02,380 --> 00:59:03,840
even into Zapier or other tools.

824
00:59:03,840 --> 00:59:08,040
We'd love to hear about those as well, so please do get in touch with us.

825
00:59:08,040 --> 00:59:11,400
Anything to share before we sign out, Martin?

826
00:59:11,400 --> 00:59:13,200
Just the email address, actually.

827
00:59:13,200 --> 00:59:19,760
People can email us their correspondence, hello at artificiallyintelligentmarketing.com.

828
00:59:19,760 --> 00:59:21,840
And a human will respond.

829
00:59:21,840 --> 00:59:22,840
Maybe.

830
00:59:22,840 --> 00:59:30,360
It might just have to be as Zapier into action and we'll set up a really complex bot to do

831
00:59:30,360 --> 00:59:31,360
that.

832
00:59:31,360 --> 00:59:34,720
Cool, thanks for sharing that, Martin, and thank you for sharing your time today.

833
00:59:34,720 --> 00:59:36,080
Lovely to hang out with you as always.

834
00:59:36,080 --> 00:59:38,160
Yes, I'm looking forward to the next one.

835
00:59:38,160 --> 00:59:39,160
Me too.

836
00:59:39,160 --> 00:59:40,160
Have a good weekend, buddy.

837
00:59:40,160 --> 00:59:41,160
Speak later.

838
00:59:41,160 --> 00:59:42,160
Bye.

839
00:59:42,160 --> 00:59:46,280
Thank you for listening to Artificially Intelligent Marketing.

840
00:59:46,280 --> 00:59:52,340
To stay on top of the latest trends, tips, and tools in the world of marketing AI, be

841
00:59:52,340 --> 00:59:54,080
sure to subscribe.

842
00:59:54,080 --> 01:00:01,680
We look forward to seeing you again next week.