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 --> 00:00:23,600
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:23,600 --> 00:00:24,680
Hello everybody.

6
00:00:24,680 --> 00:00:29,680
Welcome to Artificially Intelligent Marketing, a podcast for marketers who want to stay up

7
00:00:29,680 --> 00:00:36,040
to date on all things AI to help them improve the effectiveness and efficiency of their

8
00:00:36,040 --> 00:00:37,380
work.

9
00:00:37,380 --> 00:00:40,520
Just me today, Paul Avery, one of your co-hosts.

10
00:00:40,520 --> 00:00:41,520
Martin's not with me today.

11
00:00:41,520 --> 00:00:47,580
We've been balancing scheduling issues and Martin did a solo cast last week and this

12
00:00:47,580 --> 00:00:48,640
week.

13
00:00:48,640 --> 00:00:53,440
That sits with me, although Martin was very much kind enough to take us through his tool

14
00:00:53,440 --> 00:00:54,800
of the week this week.

15
00:00:54,800 --> 00:00:56,720
And so you will find that at the end.

16
00:00:56,720 --> 00:00:58,720
What am I going to cover today?

17
00:00:58,720 --> 00:01:04,200
Well, we're going to look at some of the news from last week, including the emergent autonomous

18
00:01:04,200 --> 00:01:07,840
scientific research capabilities of large language models.

19
00:01:07,840 --> 00:01:13,360
We're going to look at Auto GPT, which seems to be the next evolution of chat GPT and other

20
00:01:13,360 --> 00:01:17,760
tools in terms of being able to perform multiple tasks autonomously.

21
00:01:17,760 --> 00:01:22,560
We're going to look at how Hyper-Ry is T something similar, but apparently different to solve

22
00:01:22,560 --> 00:01:24,800
this similar type of problem.

23
00:01:24,800 --> 00:01:31,600
We're also going to look at AWS is announcement, Amazon's announcement for a suite of new tools

24
00:01:31,600 --> 00:01:38,480
that people can train large language models and other tools using their own data.

25
00:01:38,480 --> 00:01:44,960
So in effect, Amazon getting into the generative AI game, right?

26
00:01:44,960 --> 00:01:45,960
Let's get cracking.

27
00:01:45,960 --> 00:01:47,080
We'll get straight into it then.

28
00:01:47,080 --> 00:01:53,640
Let's talk about the this autonomous scientific research taken undertaken by a large language

29
00:01:53,640 --> 00:01:54,640
model.

30
00:01:54,640 --> 00:01:58,760
So it was a research paper released last week that showed that they were able to develop

31
00:01:58,760 --> 00:02:03,200
an intelligent, the researchers could develop an intelligent agent system that was able

32
00:02:03,200 --> 00:02:10,040
to on its own design, plan and execute scientific experiments by being given access to different

33
00:02:10,040 --> 00:02:11,320
tools.

34
00:02:11,320 --> 00:02:17,040
The system itself showed that it was capable of performing fairly complex scientific experiments.

35
00:02:17,040 --> 00:02:25,200
In one example, it was able to synthesize compounds like ibuprofen and aspirin.

36
00:02:25,200 --> 00:02:30,800
And it did that in essence by being connected to the internet and able to access a cloud

37
00:02:30,800 --> 00:02:31,800
lab.

38
00:02:31,800 --> 00:02:36,560
So a remotely accessible scientific lab, a wet lab controlled through code where the

39
00:02:36,560 --> 00:02:43,360
AI could effectively control hardware like liquid handling instruments and things like

40
00:02:43,360 --> 00:02:44,360
that.

41
00:02:44,360 --> 00:02:46,540
So what does this mean?

42
00:02:46,540 --> 00:02:52,720
It means that the researchers were able to in effect provide a brief to a large language

43
00:02:52,720 --> 00:02:56,520
model, give it access to the internet so it could gain additional information about the

44
00:02:56,520 --> 00:02:58,720
problems it was trying to solve.

45
00:02:58,720 --> 00:03:02,280
In this case, the goal was chemical synthesis.

46
00:03:02,280 --> 00:03:07,560
So they also gave it access to this virtual, well, this real life wet lab that could be

47
00:03:07,560 --> 00:03:09,300
accessed virtually.

48
00:03:09,300 --> 00:03:14,100
And the system based on its prompt could go and synthesize molecules.

49
00:03:14,100 --> 00:03:21,240
So that's kind of an exciting insight into how people are really starting to think, how

50
00:03:21,240 --> 00:03:29,000
can we use these tools to do cool stuff outside of just asking them to do research or write

51
00:03:29,000 --> 00:03:30,960
blog posts or answer questions.

52
00:03:30,960 --> 00:03:36,240
And in this case, the large language model was able to understand the input that the

53
00:03:36,240 --> 00:03:41,640
researchers provided and then go ahead by accessing things in the real world through

54
00:03:41,640 --> 00:03:47,120
APIs that connected them to this wet lab to actually synthesize chemicals.

55
00:03:47,120 --> 00:03:54,800
As you might imagine, this has pretty major implications, safety implications.

56
00:03:54,800 --> 00:04:01,360
In this case, the researchers asked the AI to construct chemicals, well-known chemicals

57
00:04:01,360 --> 00:04:04,280
that they wanted like aspirin.

58
00:04:04,280 --> 00:04:10,680
But what's to stop someone asking it to make toxic chemicals in this way?

59
00:04:10,680 --> 00:04:16,060
The other thing is, and maybe this is getting a bit Terminator style now, do we really want

60
00:04:16,060 --> 00:04:19,960
to be giving AI agents access to things in the real world like this?

61
00:04:19,960 --> 00:04:26,880
Do we really want to be giving tools the opportunity, AI to go start synthesizing things in the

62
00:04:26,880 --> 00:04:32,160
real world, in this case chemicals, but you could also imagine giving AI access to a 3D

63
00:04:32,160 --> 00:04:37,880
printer and then it could take a brief or what happens if it starts iterating and changing

64
00:04:37,880 --> 00:04:42,000
the brief and then 3D printing things in the real world as well.

65
00:04:42,000 --> 00:04:46,960
So as a proof of principle, really exciting, really cool, really interesting.

66
00:04:46,960 --> 00:04:53,400
It made Martin and I really step back and think, wow, we only had Chach-E-PT launched

67
00:04:53,400 --> 00:04:59,640
in December, GPT-4 is what, four or five weeks old and now we're looking at a large language

68
00:04:59,640 --> 00:05:05,260
model that can synthesize chemicals as kind of a little bit mind blowing really.

69
00:05:05,260 --> 00:05:12,400
In terms of what this means for marketers, perhaps in of itself, not a great deal, but

70
00:05:12,400 --> 00:05:18,840
I think it's important for us to really understand how fast language, live language models and

71
00:05:18,840 --> 00:05:25,080
their emerging capabilities are being tapped for things outside of the initial use cases

72
00:05:25,080 --> 00:05:26,080
that we saw.

73
00:05:26,080 --> 00:05:32,000
And as a part of a fairly handy segue on that topic, next thing we're going to look at is

74
00:05:32,000 --> 00:05:38,000
Auto-GPT, so the next big thing in GPT probably until next week when some other crazy thing

75
00:05:38,000 --> 00:05:39,160
happens.

76
00:05:39,160 --> 00:05:45,680
But this kind of builds on the research project I just described because in this case, Auto-GPT

77
00:05:45,680 --> 00:05:51,920
is in essence a tool that you can brief to achieve a task that requires multiple steps.

78
00:05:51,920 --> 00:05:56,040
So rather than asking a simple question or providing a fairly simple brief like we do

79
00:05:56,040 --> 00:06:04,240
to through ChatGPT for example, in essence what you could do is you could ask the tool

80
00:06:04,240 --> 00:06:10,240
to achieve an outcome and then either you can provide it with a series of actions that

81
00:06:10,240 --> 00:06:14,080
you feel it needs to go through or you can ask it what actions do you think are going

82
00:06:14,080 --> 00:06:18,080
to be required in order to achieve this goal that I've asked you to achieve.

83
00:06:18,080 --> 00:06:23,480
So to try and bring this to life, a couple of examples that have been floating around

84
00:06:23,480 --> 00:06:29,720
the Twittersphere, one user said, Auto-GPT was trying to create an app for me, recognized

85
00:06:29,720 --> 00:06:35,400
that I didn't have Node, Googled how to install Node, found a Stack Overflow article with

86
00:06:35,400 --> 00:06:40,680
the link to Node, downloaded it, extracted it and then spawned the server for me where

87
00:06:40,680 --> 00:06:43,000
all that person did was just watch.

88
00:06:43,000 --> 00:06:48,680
And this is a good example of how you didn't have to fully brief Auto-GPT on every step

89
00:06:48,680 --> 00:06:55,600
it needed to do to solve the problem, it was capable of being more resourceful than that

90
00:06:55,600 --> 00:07:00,080
to be honest and figuring out how to do it itself.

91
00:07:00,080 --> 00:07:06,920
In another example, somebody called Sully Omar provided a really nice series of tweets

92
00:07:06,920 --> 00:07:14,640
about an application for Auto-GPT that they used where Sully says, I pretended to be a

93
00:07:14,640 --> 00:07:19,880
fake shoe company and gave Auto-GPT a simple objective, do market research for waterproof

94
00:07:19,880 --> 00:07:25,920
shoes, get me the top five competitors and give me a report on their pros and cons.

95
00:07:25,920 --> 00:07:31,440
So first of all, it went straight to Google to find the top five waterproof shoe reviews.

96
00:07:31,440 --> 00:07:34,880
Once it found the links, it created questions for itself like what are the pros and cons

97
00:07:34,880 --> 00:07:40,300
of each shoe, what are the pros and cons of each top five shoe, top five waterproof shoes

98
00:07:40,300 --> 00:07:46,520
for men and started to collect and analyze this information.

99
00:07:46,520 --> 00:07:50,280
It continued to analyze the various sites, Sully says, with a combination of Googling,

100
00:07:50,280 --> 00:07:53,000
updating its queries until it was happy with the results.

101
00:07:53,000 --> 00:07:57,240
Here's an example of when it felt critically during this process, it knew that some of

102
00:07:57,240 --> 00:08:03,520
the reviews could be biased or fake, so it had to find a way to validate the reviewer.

103
00:08:03,520 --> 00:08:09,120
It even spawned its own sub-agents carrying out a task of analyzing the websites.

104
00:08:09,120 --> 00:08:15,520
There were a few times, Sully says, when it got stuck and there was no text file, it was

105
00:08:15,520 --> 00:08:19,400
able to figure out how to fix the issue all by itself.

106
00:08:19,400 --> 00:08:23,520
And the result, a pretty detailed report of the top five waterproof shoe companies with

107
00:08:23,520 --> 00:08:27,640
their pros, cons and a nice conclusion summarizing the report.

108
00:08:27,640 --> 00:08:32,080
Oh, and it only took eight minutes at a cost of 10 cents.

109
00:08:32,080 --> 00:08:38,240
So that is another fantastic example of how people are using Auto-GPT to carry out more

110
00:08:38,240 --> 00:08:46,200
complex briefs that require the agent to ask questions, make decisions, carry out its own

111
00:08:46,200 --> 00:08:50,280
research by perhaps accessing the web.

112
00:08:50,280 --> 00:08:51,880
Absolutely fascinating stuff.

113
00:08:51,880 --> 00:08:52,880
So how does this work?

114
00:08:52,880 --> 00:08:56,120
Well, it's kind of like chat GPT on steroids.

115
00:08:56,120 --> 00:08:59,880
And rather than just having a conversation, you can actually give it a brief and it takes

116
00:08:59,880 --> 00:09:01,040
a bunch of actions.

117
00:09:01,040 --> 00:09:07,980
And it does this by a series of sort of upgrades to GPT.

118
00:09:07,980 --> 00:09:13,880
So it has access to short and long term memory to be able to remember some of the outcomes

119
00:09:13,880 --> 00:09:15,480
of some of the tasks that it took.

120
00:09:15,480 --> 00:09:19,360
Obviously, it's got access to popular websites and platforms.

121
00:09:19,360 --> 00:09:21,280
It can search Google.

122
00:09:21,280 --> 00:09:27,320
It's also got access to the pine cone API to give it more amounts of vector based memory.

123
00:09:27,320 --> 00:09:30,080
It has speech to text capabilities.

124
00:09:30,080 --> 00:09:32,280
It can do image generation.

125
00:09:32,280 --> 00:09:38,320
So this is how this is much more powerful and clever than than just GPT.

126
00:09:38,320 --> 00:09:41,680
I've been trying to play with it and I have not been able to get great results yet.

127
00:09:41,680 --> 00:09:47,240
It sort of gets stuck quite easily in terms of my control of it.

128
00:09:47,240 --> 00:09:53,320
And I also have to say I haven't been using Auto-GPT itself because it's a little bit

129
00:09:53,320 --> 00:09:55,600
tricky for non-marketers to install.

130
00:09:55,600 --> 00:10:01,160
You need to visit GitHub, install the necessary files, access it via command prompt.

131
00:10:01,160 --> 00:10:05,640
I don't think it's super hard if you follow some of the online tutorials about how to

132
00:10:05,640 --> 00:10:09,560
do this, but it might be hard enough for those of us that think they see the word Python

133
00:10:09,560 --> 00:10:14,640
and get chills down our spine or we're like, oh my goodness, I have to open up the terminal

134
00:10:14,640 --> 00:10:15,640
and start running commands.

135
00:10:15,640 --> 00:10:19,160
I'm not sure how keen I am to get into that either.

136
00:10:19,160 --> 00:10:23,360
But that's why I was using some browser based implementations of this and several have popped

137
00:10:23,360 --> 00:10:24,360
up.

138
00:10:24,360 --> 00:10:28,560
So there's Cognosys.ai and Godmode.space.

139
00:10:28,560 --> 00:10:34,640
And in essence, they have brought, at least as far as I can tell, a decent chunk of Auto-GPT's

140
00:10:34,640 --> 00:10:40,960
capabilities to a browser based interface where you can drop your prompt in and then

141
00:10:40,960 --> 00:10:46,440
basically run Auto-GPT without having to install it all on your machine.

142
00:10:46,440 --> 00:10:50,660
So I would recommend that people who are interested in this stuff and trying to think about how

143
00:10:50,660 --> 00:10:55,900
they could give AI more complex briefs to help them achieve more complex projects actually

144
00:10:55,900 --> 00:10:57,800
have a play with this.

145
00:10:57,800 --> 00:11:01,500
By drifting ever closer to an assistant that can do more than just answer simple questions

146
00:11:01,500 --> 00:11:04,320
or produce short from content like blog posts.

147
00:11:04,320 --> 00:11:07,280
Bunch of examples I've given already, but a few other things have been flown around the

148
00:11:07,280 --> 00:11:12,320
web is to have it do market research and create a business plan, ask it to gather data and

149
00:11:12,320 --> 00:11:17,320
stats from the web and summarise as a report, find and compare products from various websites

150
00:11:17,320 --> 00:11:21,320
as sort of akin to one of the examples I've just recently gave.

151
00:11:21,320 --> 00:11:26,080
And also, alternating customer service inquiries and responses by giving the tool access to

152
00:11:26,080 --> 00:11:28,280
data about products, etc.

153
00:11:28,280 --> 00:11:30,720
So yes, a very interesting development.

154
00:11:30,720 --> 00:11:36,520
And again, all of this has happened since GPT-4 was launched about four weeks ago and

155
00:11:36,520 --> 00:11:38,920
TrackGPT was just before Christmas.

156
00:11:38,920 --> 00:11:41,320
The speed here is insane.

157
00:11:41,320 --> 00:11:46,520
Now there is one word of warning for those of you that want to go run off now and start

158
00:11:46,520 --> 00:11:47,520
playing with Auto-GPT.

159
00:11:47,520 --> 00:11:52,020
Certainly the browser based tools that I was playing with, I think they got overloaded

160
00:11:52,020 --> 00:11:53,020
pretty quickly.

161
00:11:53,020 --> 00:11:57,200
It's probably costing them a fair bit of money and so now if you want to access them, you

162
00:11:57,200 --> 00:12:01,640
have to provide your own OpenAI API key.

163
00:12:01,640 --> 00:12:04,560
Try and say that after a few runs.

164
00:12:04,560 --> 00:12:07,780
In essence, what this means, you have to log into your OpenAI account, you have to go to

165
00:12:07,780 --> 00:12:12,220
where you can generate your own API keys, which is like a string of numbers and letters

166
00:12:12,220 --> 00:12:17,240
that's unique to you that allows you to access the system almost through the back end, if

167
00:12:17,240 --> 00:12:18,560
you like.

168
00:12:18,560 --> 00:12:21,520
And you have to then paste that key into one of these tools.

169
00:12:21,520 --> 00:12:24,320
There's a certain amount of trust that comes with doing that.

170
00:12:24,320 --> 00:12:28,560
Plus accessing the OpenAI API is something you have to pay for.

171
00:12:28,560 --> 00:12:33,920
So you'll have to put some cash on your credit within OpenAI.

172
00:12:33,920 --> 00:12:39,300
I have heard that the system can absolutely eat OpenAI credits for fun.

173
00:12:39,300 --> 00:12:41,460
So just be prepared to keep an eye on that.

174
00:12:41,460 --> 00:12:46,560
You might find you burn through cash a bit faster than you might have expected.

175
00:12:46,560 --> 00:12:48,680
I don't think we're talking thousands of dollars here.

176
00:12:48,680 --> 00:12:54,600
Plus, you have to put credit on there like an old school mobile cell phone where you'd

177
00:12:54,600 --> 00:12:57,120
put 10 dollars of credit, 10 pounds of credit on your phone.

178
00:12:57,120 --> 00:12:58,320
And when it was gone, it was gone.

179
00:12:58,320 --> 00:13:01,880
So I think as long as you set it up like that, you're not going to end up going, crumbs,

180
00:13:01,880 --> 00:13:04,840
I've just spent $5,000 on this, which I think would be pretty hard anyway.

181
00:13:04,840 --> 00:13:10,980
But just something to keep in mind if you plan to try using these systems.

182
00:13:10,980 --> 00:13:17,080
On a related note, for those of you that use generative AI writing tools, one of the interesting

183
00:13:17,080 --> 00:13:20,080
tools in the space is called HyperWrite.

184
00:13:20,080 --> 00:13:25,880
And HyperWrite team announced on Twitter this week that they're testing an AI agent that

185
00:13:25,880 --> 00:13:27,920
can use the internet like a human.

186
00:13:27,920 --> 00:13:32,580
So in the example that they show, it orders a pizza from Domino's with a single command.

187
00:13:32,580 --> 00:13:34,600
And it's like a Chrome plugin.

188
00:13:34,600 --> 00:13:39,960
So it just basically browsers the web and interacts with websites based on the command

189
00:13:39,960 --> 00:13:41,300
that you give it.

190
00:13:41,300 --> 00:13:45,400
So if you are a user of HyperWrite, you can get on the waiting list for this because they're

191
00:13:45,400 --> 00:13:48,080
providing early access to some of their users.

192
00:13:48,080 --> 00:13:54,720
So just log into your account and ask to be added to the wait list or create account and

193
00:13:54,720 --> 00:13:57,200
ask to be added to the wait list.

194
00:13:57,200 --> 00:14:01,300
The system promises that if you just describe what you want it to do, it will automatically

195
00:14:01,300 --> 00:14:03,840
operate Chrome for you to achieve your task.

196
00:14:03,840 --> 00:14:06,860
I have not had a chance to play with this yet.

197
00:14:06,860 --> 00:14:11,520
So I do not know how buggy or powerful it is, but you could imagine that it could be

198
00:14:11,520 --> 00:14:17,600
really powerful for automating tasks that are more complex where you have to be able

199
00:14:17,600 --> 00:14:20,800
to browse different websites.

200
00:14:20,800 --> 00:14:26,900
The market research examples from before spring to mind, not that I'm advocating it, but also

201
00:14:26,900 --> 00:14:33,120
the potential to scrape content from the pages of many websites and then interrogate that

202
00:14:33,120 --> 00:14:38,440
content through large, through natural language questions could be another example.

203
00:14:38,440 --> 00:14:43,780
The team on Twitter said that they could imagine anyone doing anything from booking flights

204
00:14:43,780 --> 00:14:49,400
to ordering food to researching complex topics and having the system manage their email.

205
00:14:49,400 --> 00:14:54,320
So when we get a chance to play with it, we will report back on how good or not good it

206
00:14:54,320 --> 00:14:59,640
is compared to how good it sounds, which it certainly sounds really, really good.

207
00:14:59,640 --> 00:15:02,760
For marketers, I think get yourself on the waiting list and have a play because this

208
00:15:02,760 --> 00:15:04,800
is another one of those things.

209
00:15:04,800 --> 00:15:08,800
Could you automate with the system staying abreast of what your competitors are doing

210
00:15:08,800 --> 00:15:12,680
by monitoring their websites and when they put a new product up on the website, perhaps

211
00:15:12,680 --> 00:15:17,360
there'll be some mechanism by which you can trigger an alert or have it added to a spreadsheet

212
00:15:17,360 --> 00:15:18,360
or something.

213
00:15:18,360 --> 00:15:20,680
That's not clear to me yet, but I think that would be pretty cool.

214
00:15:20,680 --> 00:15:25,440
You could do the same thing for things like conducting brand sentiment analysis, maybe

215
00:15:25,440 --> 00:15:27,960
doing clever things with Google Alerts, who knows?

216
00:15:27,960 --> 00:15:33,920
So I think one to watch and to have a play with if you get a chance.

217
00:15:33,920 --> 00:15:37,720
Next story, we're going to talk very briefly about Amazon.

218
00:15:37,720 --> 00:15:43,000
So Amazon, where have you been during the whole generative AI and wider AI explosion

219
00:15:43,000 --> 00:15:44,240
of the last few months?

220
00:15:44,240 --> 00:15:50,080
Well, the answer would appear to be they have been busily buzzing in the background.

221
00:15:50,080 --> 00:15:53,240
Another thing that would probably be hard to say after a color arms.

222
00:15:53,240 --> 00:15:58,120
They have released their Amazon Bedrock, which is a service that provides access to what

223
00:15:58,120 --> 00:16:04,360
they're calling foundational models for generative AI, which includes models from AI21 Labs,

224
00:16:04,360 --> 00:16:12,080
Anthropic, Stability AI and Amazon slipped their Titan model in there at the same time.

225
00:16:12,080 --> 00:16:17,240
Bedrock, they are claiming, is a scalable, reliable, secure AWS managed service where

226
00:16:17,240 --> 00:16:23,800
customers can customize models by pointing Bedrock to label examples within the Amazon

227
00:16:23,800 --> 00:16:24,800
cloud.

228
00:16:24,800 --> 00:16:34,280
So in essence, if you are using Amazon's AWS or you're using cloud tools from Amazon and

229
00:16:34,280 --> 00:16:39,560
that you're hosting data there, one assumes that you will be able to leverage Bedrock,

230
00:16:39,560 --> 00:16:46,800
especially if your data is somewhat labeled or organized in some way to, in essence, create

231
00:16:46,800 --> 00:16:57,120
your own chat bots or other large language model driven tools that lean on the large

232
00:16:57,120 --> 00:17:02,480
language, natural language processing abilities that come from these large models, but that

233
00:17:02,480 --> 00:17:05,280
is trained on your own data.

234
00:17:05,280 --> 00:17:10,520
So something that would instantly spring to mind for me is something like to be able to

235
00:17:10,520 --> 00:17:15,880
build your own knowledge base for your website that was a chat bot instead of a bunch of

236
00:17:15,880 --> 00:17:21,640
web pages, because you could train, you could train Bedrock based on that.

237
00:17:21,640 --> 00:17:25,680
One assumes that it wouldn't be that hard to train Bedrock on data that wasn't already

238
00:17:25,680 --> 00:17:29,240
on AWS cloud system as well.

239
00:17:29,240 --> 00:17:32,560
But to be honest, I have to dive a bit deeper into that to really figure out what we can

240
00:17:32,560 --> 00:17:33,560
do.

241
00:17:33,560 --> 00:17:40,840
But it's certainly worth looking at if you are interested in creating LLM based tools

242
00:17:40,840 --> 00:17:45,840
like chat bots, but that you want there to be able to answer specific questions on your

243
00:17:45,840 --> 00:17:47,960
own data, information and content.

244
00:17:47,960 --> 00:17:52,440
For example, if you've got a bunch of products and you're an e-commerce store, or if you've

245
00:17:52,440 --> 00:17:57,360
got a large knowledge base or lots of technical documentation that you want to be able to

246
00:17:57,360 --> 00:18:01,520
provide easier access to your customers in a way that they could just ask it through

247
00:18:01,520 --> 00:18:04,520
natural language.

248
00:18:04,520 --> 00:18:09,000
Martin was the first one to bring this to light, shared it with me on the Amazons and

249
00:18:09,000 --> 00:18:13,360
he's sent over a few thoughts here.

250
00:18:13,360 --> 00:18:18,200
This is a really exciting marketplace for foundational models that developers building

251
00:18:18,200 --> 00:18:22,360
on AWS can now easily choose to integrate some of the best foundational models on the

252
00:18:22,360 --> 00:18:26,040
market into their existing applications that they've built on AWS.

253
00:18:26,040 --> 00:18:30,720
Because of course, not just websites and things that live on AWS, it's also apps and other

254
00:18:30,720 --> 00:18:34,520
software tools.

255
00:18:34,520 --> 00:18:40,480
I don't think Martin thinks that Titan as Amazon's own model is going to contribute

256
00:18:40,480 --> 00:18:42,320
anything above what we've already seen.

257
00:18:42,320 --> 00:18:47,640
I think it's more the ease of access to a range of different models that's exciting

258
00:18:47,640 --> 00:18:52,560
here.

259
00:18:52,560 --> 00:19:00,840
When he looks at how this is going to emerge and develop, he's not sure that Amazon are

260
00:19:00,840 --> 00:19:08,640
as confident in their offering as open AI and Microsoft to a forging really far ahead.

261
00:19:08,640 --> 00:19:13,480
So I think Amazon's move here is about time probably.

262
00:19:13,480 --> 00:19:15,560
It's going to be one to watch.

263
00:19:15,560 --> 00:19:17,880
We don't get the feeling that they're on the cutting edge here.

264
00:19:17,880 --> 00:19:22,240
They're probably trailing, especially because mostly they're clear to be leveraging the

265
00:19:22,240 --> 00:19:26,440
power of other people's models.

266
00:19:26,440 --> 00:19:34,280
Right, last story for this week, which is that Google CEO Sundar Pichai warned society

267
00:19:34,280 --> 00:19:37,280
to brace for the impact of AI acceleration.

268
00:19:37,280 --> 00:19:44,360
So Sundar was on an interview with CBS's 60 Minutes on Sunday.

269
00:19:44,360 --> 00:19:49,800
And in it, he hinted that society isn't prepared for the rapid advancement of AI.

270
00:19:49,800 --> 00:19:56,000
Warning of the consequences, he said AI will impact every product of every company.

271
00:19:56,000 --> 00:20:01,840
Little bit of a pause just to let that set in there because I think that is the scale

272
00:20:01,840 --> 00:20:08,080
of how we have to really start to think about this as marketers and as business people.

273
00:20:08,080 --> 00:20:11,860
When you think about some of the things we've featured on the podcast recently, we spoke

274
00:20:11,860 --> 00:20:18,320
a few weeks ago about how a large collection of tech experts called for a six month AI

275
00:20:18,320 --> 00:20:26,160
and hiatus, you know, stop launching and bringing to market new AI developments.

276
00:20:26,160 --> 00:20:30,160
We also looked at OpenAI's research paper on the impact of AI on the labor market, where

277
00:20:30,160 --> 00:20:35,240
they predicted that as many as one fifth of the workforce could have half of their tasks

278
00:20:35,240 --> 00:20:37,320
affected by AI.

279
00:20:37,320 --> 00:20:42,700
And so what we see from here is that Sundar is continuing on this theme, accentuating

280
00:20:42,700 --> 00:20:48,280
the need for society to carefully consider and plan for the imminent impact of AI on

281
00:20:48,280 --> 00:20:49,280
society.

282
00:20:49,280 --> 00:20:54,920
So the take home message here for us is that people in the know continue to have serious

283
00:20:54,920 --> 00:21:01,440
concerns about the impact of AI on society, especially the labor market.

284
00:21:01,440 --> 00:21:04,160
Now what does this mean for marketers?

285
00:21:04,160 --> 00:21:09,640
Our thoughts here on the podcast are that the genie's out of the bottle on this one.

286
00:21:09,640 --> 00:21:13,720
We don't see this slowing down anytime soon, as we've mentioned on previous podcasts.

287
00:21:13,720 --> 00:21:19,040
And as a marketer, we feel for all of us, our best recourse is to stay on top of the

288
00:21:19,040 --> 00:21:24,680
topic, play with the tools as they become available and become an expert on using AI

289
00:21:24,680 --> 00:21:26,960
to augment your work.

290
00:21:26,960 --> 00:21:29,680
And it really is still augmentation, right?

291
00:21:29,680 --> 00:21:32,040
Many of the tools are very powerful.

292
00:21:32,040 --> 00:21:37,820
We've talked a little bit about AutoGPT today, but they are still massively capable of hallucination

293
00:21:37,820 --> 00:21:39,400
and producing junk.

294
00:21:39,400 --> 00:21:43,880
There's loads of examples where AutoGPT just gets stuck.

295
00:21:43,880 --> 00:21:48,240
So I really do think human in the loop is still going to be critical.

296
00:21:48,240 --> 00:21:51,880
And where we're going to add the value here is by augmenting ourselves.

297
00:21:51,880 --> 00:21:56,480
So we need to learn to get really good at prompting these systems, thinking of smart

298
00:21:56,480 --> 00:21:58,360
things that we want done.

299
00:21:58,360 --> 00:22:00,200
So there's the creativity aspect to that.

300
00:22:00,200 --> 00:22:04,520
And then learning how to prompt the system to actually get what we want.

301
00:22:04,520 --> 00:22:08,720
And then I think it's about bringing your expertise to double check the outputs you

302
00:22:08,720 --> 00:22:15,360
get to make sure it's not full of errors, lies, hallucinations, the type of thing that

303
00:22:15,360 --> 00:22:20,280
will be mistakes that we just, none of us can afford to have in reports or content or

304
00:22:20,280 --> 00:22:22,320
anything that we produce.

305
00:22:22,320 --> 00:22:25,360
There's also the copyright aspect that we need to keep an eye on here as well.

306
00:22:25,360 --> 00:22:32,320
Of course, we've talked about on previous podcasts, but I do think the way to resist

307
00:22:32,320 --> 00:22:37,160
the challenges here is to figure out how you're going to use AI to augment yourself and that

308
00:22:37,160 --> 00:22:42,720
this AI human collaboration is where the magic is going to be.

309
00:22:42,720 --> 00:22:47,320
So whilst there's quite a bit of fear mongering to this, and there's been quite a bit of debate

310
00:22:47,320 --> 00:22:53,520
on Twitter and LinkedIn about the fear mongering aspects, just practically speaking, these

311
00:22:53,520 --> 00:22:58,280
tools can probably help you be more effective, your team be more effective.

312
00:22:58,280 --> 00:23:03,440
And I think your job and our job in marketing is to figure out how we can leverage them

313
00:23:03,440 --> 00:23:09,280
to maximise our impact rather than perhaps fearing they're going to take our jobs.

314
00:23:09,280 --> 00:23:14,160
How are we going to work with them to be even more efficient and effective in what we're

315
00:23:14,160 --> 00:23:15,560
doing?

316
00:23:15,560 --> 00:23:20,120
So that is artificially intelligent marketing news for this week.

317
00:23:20,120 --> 00:23:22,840
From here, we're going to transition into tool of the week.

318
00:23:22,840 --> 00:23:26,440
And then after that, we should be back to normal, hopefully next week.

319
00:23:26,440 --> 00:23:32,120
And apologies for the lack of witty or less than witty banter, but next week, I promise

320
00:23:32,120 --> 00:23:38,160
you we'll be back talking about how poor Derby County are, how depressed Martin is about

321
00:23:38,160 --> 00:23:39,160
it.

322
00:23:39,160 --> 00:23:43,720
And I will continue to tease him mercilessly for those that have been enjoying that part

323
00:23:43,720 --> 00:23:44,720
of the podcast.

324
00:23:44,720 --> 00:23:45,720
Right.

325
00:23:45,720 --> 00:23:48,520
We look forward to getting back into our Northman Cairns next week.

326
00:23:48,520 --> 00:23:52,280
If you've enjoyed the podcast, please subscribe, share it with other marketers in your network

327
00:23:52,280 --> 00:23:53,280
if you feel they benefit.

328
00:23:53,280 --> 00:23:56,440
We look forward to seeing you all next week.

329
00:23:56,440 --> 00:24:04,760
This week's tool of the week is ClipDrop brought to us by Stable AI, the company who created

330
00:24:04,760 --> 00:24:09,560
Stable Diffusion, the open source text to image generation tool that many of us will

331
00:24:09,560 --> 00:24:10,560
have been using.

332
00:24:10,560 --> 00:24:19,440
Now ClipDrop is available through cliptrop.co and it's an online interface that's very

333
00:24:19,440 --> 00:24:26,520
easy to use and it's all powered by the Stable Diffusion image generation model.

334
00:24:26,520 --> 00:24:30,200
So what makes it different from the Stable Diffusion model that many of us will have

335
00:24:30,200 --> 00:24:35,440
been using because we cannot actually use Stable Diffusion in loads of tools already.

336
00:24:35,440 --> 00:24:42,200
For example, Canva's own text to image generation tool is in fact powered by Stable Diffusion.

337
00:24:42,200 --> 00:24:47,960
Well, what makes ClipDrop different is the fact that they've put together a handful of

338
00:24:47,960 --> 00:24:55,480
tools, in fact about nine or ten tools they've put together that you can do more than just

339
00:24:55,480 --> 00:24:58,000
generate images from text with.

340
00:24:58,000 --> 00:25:03,840
Of course you can generate high resolution images with a simple text description, but

341
00:25:03,840 --> 00:25:09,000
they've really taken the power of Stable Diffusion and given it some very specific

342
00:25:09,000 --> 00:25:10,000
applications.

343
00:25:10,000 --> 00:25:12,720
For example, you can clean up images.

344
00:25:12,720 --> 00:25:17,280
You can upload a photo, maybe it's a photo with a coffee cup in the foreground that you

345
00:25:17,280 --> 00:25:22,240
want to get rid of and just by masking it off you can say clean up and then it will

346
00:25:22,240 --> 00:25:29,280
get rid of that image, blending it perfectly into the rest of the image.

347
00:25:29,280 --> 00:25:34,640
Very similar to Google Photos magic eraser feature.

348
00:25:34,640 --> 00:25:39,800
You've also got the remove background which is pretty much as you would expect.

349
00:25:39,800 --> 00:25:44,560
It takes out the background from an image and gives you a very neat cutout of the object

350
00:25:44,560 --> 00:25:50,680
that you wanted to get, very similar to what you'll have seen in Canva already and these

351
00:25:50,680 --> 00:25:55,680
tools are available online as standalone features themselves.

352
00:25:55,680 --> 00:26:03,600
I think they don't have much of a future based on how readily available these AI powered

353
00:26:03,600 --> 00:26:07,040
cutout tools are these days.

354
00:26:07,040 --> 00:26:11,960
There's also a relight feature which is quite a neat one I think.

355
00:26:11,960 --> 00:26:18,280
It allows you to basically add a lighting studio to photos that you've already taken.

356
00:26:18,280 --> 00:26:23,480
You can relight your images, maybe you've taken a portrait photo of someone and you

357
00:26:23,480 --> 00:26:28,280
want to add a bit of side lighting that you didn't have the opportunity to do when you

358
00:26:28,280 --> 00:26:32,480
took the photo and you can do that and it works really well.

359
00:26:32,480 --> 00:26:40,440
But one of the tools I think is particularly useful for say marketing agencies is the text

360
00:26:40,440 --> 00:26:44,720
remover which will remove text from any image.

361
00:26:44,720 --> 00:26:49,600
So if you've got a billboard that you've got a photo of and you think it looks great and

362
00:26:49,600 --> 00:26:55,780
you could maybe drop in a client mockup on this billboard but it's got some text on it,

363
00:26:55,780 --> 00:27:00,520
you can just upload the photo into ClipDrop and at the touch of a button it will remove

364
00:27:00,520 --> 00:27:06,640
all of the text from that image giving you a nice clean space where you can then add

365
00:27:06,640 --> 00:27:11,000
your own campaign graphics straight into the image.

366
00:27:11,000 --> 00:27:15,320
The other thing you can do is the replace background tool and this is where it combines

367
00:27:15,320 --> 00:27:21,680
the text generation that we've seen through stable diffusion with a new capability.

368
00:27:21,680 --> 00:27:26,560
So you will add in for instance a product photo, maybe it's a, the example they give

369
00:27:26,560 --> 00:27:29,040
on the website is a bottle of wine.

370
00:27:29,040 --> 00:27:38,840
You add in your photo and then you want to type in a hosier marble kitchen and it will

371
00:27:38,840 --> 00:27:46,720
drop that image straight into, drop that bottle of wine straight into a marble kitchen, marble

372
00:27:46,720 --> 00:27:49,080
work tops and what have you.

373
00:27:49,080 --> 00:27:57,280
So for creating product mockups, great for e-commerce, it can be implemented really quickly.

374
00:27:57,280 --> 00:28:02,240
If you've got a photo that you like but you're not quite happy with it, maybe you want to

375
00:28:02,240 --> 00:28:09,080
see a few variations, they've created the stable diffusion re-imagine tool where you

376
00:28:09,080 --> 00:28:15,600
can upload an existing image and it will basically remix new images that are very similar.

377
00:28:15,600 --> 00:28:19,080
Now looking through the website I actually thought it was quite interesting some of the

378
00:28:19,080 --> 00:28:26,480
examples that they provided because there is one example which has like a living room

379
00:28:26,480 --> 00:28:33,040
scene with sofas and chairs in the original image and lots of coffee tables and then it

380
00:28:33,040 --> 00:28:35,240
shows you some remixed versions.

381
00:28:35,240 --> 00:28:40,840
Now I must say that the remixed versions that it shows in the demo are not great because

382
00:28:40,840 --> 00:28:45,280
you just have to look at them for a couple of seconds to realize that the chair legs

383
00:28:45,280 --> 00:28:53,640
have a few too many legs and the table and the chairs seem to be merged into one another.

384
00:28:53,640 --> 00:28:59,200
So at least Stability AI are kind of showing off the reality of working with these models,

385
00:28:59,200 --> 00:29:06,040
these image generation tools are not perfect, they still struggle with counting chair legs

386
00:29:06,040 --> 00:29:13,280
and making sure that furniture would realistically balance in the real world and of course they

387
00:29:13,280 --> 00:29:19,280
still struggle with the number of fingers that people have on their hands.

388
00:29:19,280 --> 00:29:23,120
But overall I think ClipDrop is a pretty neat tool to play around with.

389
00:29:23,120 --> 00:29:27,720
If you've never played with the Text2Image tools then this is one of the better ones

390
00:29:27,720 --> 00:29:30,080
out there to get started with for free.

391
00:29:30,080 --> 00:29:33,560
You can have a play with it, the pricing is pretty good.

392
00:29:33,560 --> 00:29:40,880
Like I say you can get started for free, there is a maximum size image that you can work

393
00:29:40,880 --> 00:29:49,520
with at the free tier, for instance the background removal you can work up to I think it's 1024

394
00:29:49,520 --> 00:30:00,640
by 1024 max so fairly small images and the image upscaler is a 2x upscaler.

395
00:30:00,640 --> 00:30:07,400
There is a Android and iOS app, although I haven't tried either of those at this stage,

396
00:30:07,400 --> 00:30:12,080
I was just playing around with the web interface at this point.

397
00:30:12,080 --> 00:30:17,960
If you want faster generations and you want unlimited capability, so unlimited background

398
00:30:17,960 --> 00:30:25,960
removal, if you want unlimited clean up pictures and web editing and what have you, it's £5

399
00:30:25,960 --> 00:30:28,600
a month, so it's really competitively priced.

400
00:30:28,600 --> 00:30:33,840
I think that's an incredibly good plan.

401
00:30:33,840 --> 00:30:37,680
If you are a developer and you're looking to integrate stable diffusion and some of

402
00:30:37,680 --> 00:30:44,200
these capabilities into your product it does have an API as well.

403
00:30:44,200 --> 00:30:50,880
I think really what the ability AI have done with ClipDrop is made the stable diffusion

404
00:30:50,880 --> 00:30:57,400
model very very accessible to everyday users.

405
00:30:57,400 --> 00:31:04,920
Rather than having to subscribe to some obscure tool or try to hack something together using

406
00:31:04,920 --> 00:31:11,240
an API if you're not a developer or even because it is open source you can download the model

407
00:31:11,240 --> 00:31:15,240
and install it and run it on your own servers.

408
00:31:15,240 --> 00:31:20,360
You don't have to do any of that, you can just use this web interface and it feels like

409
00:31:20,360 --> 00:31:27,400
this is a very easy way for people to get into using AI powered image generation and

410
00:31:27,400 --> 00:31:28,400
image editing.

411
00:31:28,400 --> 00:31:32,360
So yep, that's my tool of the week ClipDrop.co.

412
00:31:32,360 --> 00:31:37,240
Go have a play with it if you're interested in image generation and image editing and

413
00:31:37,240 --> 00:31:43,040
you're interested in seeing where stable diffusion is likely to go next because clearly

414
00:31:43,040 --> 00:31:52,560
Stability AI are putting a lot of focus into this foundational model.

415
00:31:52,560 --> 00:31:56,260
Thank you for listening to Artificially Intelligent Marketing.

416
00:31:56,260 --> 00:32:02,320
To stay on top of the latest trends, tips and tools in the world of marketing AI be

417
00:32:02,320 --> 00:32:04,080
sure to subscribe.

418
00:32:04,080 --> 00:32:07,640
We look forward to seeing you again next week.