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

2
00:00:10,000 --> 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 --> 00:00:23,040
results from your marketing efforts. Now let's join our hosts, Paul Avery and Martin Broadhurst.

4
00:00:23,040 --> 00:00:28,840
Hello everyone, welcome to Episode 40 of Artificially Intelligent Marketing. It's Paul Avery here

5
00:00:28,840 --> 00:00:35,080
and I'm joined as always by Martin Broadhurst, the fantabulous AI guru. Martin, how are you, pal?

6
00:00:35,080 --> 00:00:41,240
Overwhelmed with new AI tools to play with. You have been having a tinker this week and don't worry,

7
00:00:41,240 --> 00:00:44,520
dear listener, we're going to get into that tinkering in some detail as we go through the

8
00:00:44,520 --> 00:00:48,680
stories today. Why don't we just get straight into it, Martin? There's no messing about it.

9
00:00:48,680 --> 00:00:53,320
I wanted to talk to Arby County briefly because apparently there's a team in Spain that's about

10
00:00:53,320 --> 00:00:58,120
to have the worst league finish ever or at least they're on course for it. And of course,

11
00:00:58,120 --> 00:01:02,440
your team, Derby County, currently hold that record so you must be pretty chuffed.

12
00:01:02,440 --> 00:01:07,240
Currently the worst team in history with 11 points from an entire season, one victory.

13
00:01:07,240 --> 00:01:13,240
It was bleak that year, I remember it well. So the idea that somebody's going to take that record

14
00:01:13,240 --> 00:01:20,120
from us is welcome because we've had that since 2008. And to be honest, I didn't think it would

15
00:01:20,120 --> 00:01:25,560
ever be beaten. I think there's still time for them to turn it around but I got my fingers crossed

16
00:01:25,560 --> 00:01:31,960
for you. Anyway, that's digression. Let's get into the AI. So we've been on the air for a couple of

17
00:01:31,960 --> 00:01:37,400
weeks. Some interesting stuff has happened. Probably the most exciting was the release of

18
00:01:37,400 --> 00:01:44,920
Gemini Ultra from our good friends over at Google. If you are a casual listener to the podcast or a

19
00:01:45,640 --> 00:01:50,680
casual person paying attention to AI, you could be forgiven for trying to figure out what are all

20
00:01:50,680 --> 00:01:56,040
the different names of all the things that keep changing their name. So this was called Bard.

21
00:01:56,600 --> 00:02:02,920
It's now called Gemini. Ultra is like Google's awesome most powerful model that's supposedly

22
00:02:02,920 --> 00:02:09,480
on the level of GPT-4. And you can access it as a chat bot, but also it's going to be in start

23
00:02:09,480 --> 00:02:15,560
baked into Google's products like Gmail and Docs as part of what was called Google Duet and may

24
00:02:15,560 --> 00:02:21,480
still be called Google Duet at least for the next 15 minutes or so, who knows. But for the purposes

25
00:02:21,480 --> 00:02:28,680
of this conversation, Google have released their new most powerful version of chat GPT-4. This is

26
00:02:28,680 --> 00:02:34,040
the closest that we've got to a GPT-4 rival and that's why it's interesting and exciting.

27
00:02:34,040 --> 00:02:39,880
A, because now we've got another GPT-4 level tool we can play with. And B, because as we've

28
00:02:39,880 --> 00:02:44,120
talked about in the podcast before, different providers innovating in different ways with their

29
00:02:44,120 --> 00:02:49,400
tools stimulates each other to keep finding new routes forward to give us even better tools. So

30
00:02:49,400 --> 00:02:57,640
GPT-4.5 or 5 can now be much closer potentially because of this stimulation that Gemini Ultra

31
00:02:57,640 --> 00:03:02,120
is going to bring to the market. But yes, it's available. We've been playing with it. I know

32
00:03:02,120 --> 00:03:07,000
Martin, you got straight in, got yourself a paid subscription. We should all say many thanks to

33
00:03:07,000 --> 00:03:11,240
Martin. He buys a lot of subscriptions to play with all these tools so that he can tell us if

34
00:03:11,240 --> 00:03:19,720
they're any good or not. Tell us mine. Is Gemini Ultra any good? In short, yes. The slightly longer

35
00:03:19,720 --> 00:03:27,320
answer requires a slightly different question, I feel. Is it as good as GPT-4? And I think we'll

36
00:03:28,680 --> 00:03:34,200
get there is what I'll say. First and foremost, it is a very good model and it does feel very

37
00:03:34,200 --> 00:03:42,120
different from Gemini Pro. For anyone that isn't aware of the pricing model of this, it's £18.99

38
00:03:43,960 --> 00:03:50,360
British pounds per month. And it comes as part of an upgraded Google One plan. If you're somebody

39
00:03:50,360 --> 00:03:56,680
who's been using Google One to get the increased storage space on Gmail and Google Photos and all

40
00:03:56,680 --> 00:04:02,600
of that kind of thing, this is an upgrade where you can actually get two terabytes of storage,

41
00:04:02,600 --> 00:04:09,480
as well as access to Gemini Advanced and Ultra. But that's about as much as the extra perks go.

42
00:04:10,280 --> 00:04:18,120
So it's cheaper than chat GPT with GPT-4. And you get these additional bonuses of the increased

43
00:04:18,120 --> 00:04:25,720
cloud storage and what have you. Compared to Gemini Pro, it's an immediate step up. You just

44
00:04:25,720 --> 00:04:31,400
feel that. I think when Gemini Pro was launched pre-Christmas, we all tried it and said, yeah,

45
00:04:31,400 --> 00:04:40,760
this feels like a GPT-3.5 capable model. And it's free, right? So Gemini Pro was free.

46
00:04:40,760 --> 00:04:44,120
If you had a Google account, you could go play with it. Even Google Workspace, I think you could

47
00:04:44,120 --> 00:04:50,440
go play with it. So to have a free model that was at GPT-3.5 or maybe even just above for some stuff

48
00:04:51,240 --> 00:04:56,600
was nice. But yes, compared to a paid model, we're expecting a leap, I would say.

49
00:04:56,600 --> 00:05:03,880
Yeah. And this definitely is a leap. One word of note is that it's only available on personal

50
00:05:04,840 --> 00:05:11,560
Workspace accounts or Google One account. It's not available through business accounts,

51
00:05:11,560 --> 00:05:22,600
which is slightly frustrating. And at the moment, that means that any input is not private, right?

52
00:05:22,600 --> 00:05:28,520
So in the same way that with a Google Plus account, sorry, not Google Plus, Google Plus,

53
00:05:28,520 --> 00:05:34,120
that's going back a few years. Bring it back. AI-driven Google Plus, bring it back.

54
00:05:37,160 --> 00:05:44,040
The chat GPT accounts where you have to be careful what you're putting in because the data could be

55
00:05:45,080 --> 00:05:52,360
viewed or going into the training model. This is the case with Gemini Ultra as it stands. In fact,

56
00:05:52,360 --> 00:05:59,960
there is a warning that says prompts and inputs can be reviewed by human reviewers. So don't

57
00:05:59,960 --> 00:06:10,200
say anything that you wouldn't want another human being to read. Crumps. Yeah. In terms of the

58
00:06:10,200 --> 00:06:15,160
interface is very similar to what we're familiar with on the left-hand side. You've got your

59
00:06:15,720 --> 00:06:20,760
history of your chat and the main central column is where you have your conversations.

60
00:06:20,760 --> 00:06:30,360
Now the inputs are text-based and you can also input images. There is a microphone button

61
00:06:31,000 --> 00:06:38,840
that when I clicked it, I was fully expecting it to record audio and upload that as an audio

62
00:06:38,840 --> 00:06:45,800
input prompt because when they announced Gemini Ultra, this is a multimodal capability. And they

63
00:06:45,800 --> 00:06:51,640
said that one of the modalities that you get with it is the idea that you can input audio

64
00:06:52,680 --> 00:06:59,480
and it will be able to work with audio input. Unfortunately, if you hit record, all it does

65
00:06:59,480 --> 00:07:07,320
is transcribe the text. So it's the same as the chat GPT microphone button. It takes a spoken word,

66
00:07:07,320 --> 00:07:11,160
turns it into text and inputs it as text. That was slightly disappointing.

67
00:07:11,160 --> 00:07:19,080
Okay. So quality of the outputs. I spent a good few hours playing with it, tried it with creative

68
00:07:19,880 --> 00:07:28,040
responses, but also tried to use it as a programmer and a kind of coding assistant and came up with a

69
00:07:28,040 --> 00:07:35,640
project and wanted to jump into it and see how far I got. And overall the responses are very good.

70
00:07:35,640 --> 00:07:46,040
It feels like GPT-4 at that level. And if GPT-4 didn't exist, I would say this was the best on the

71
00:07:46,040 --> 00:07:54,680
market by a margin. However, GPT-4 does exist and this doesn't quite get there. So a few things

72
00:07:54,680 --> 00:08:04,520
that kind of let it down. When coding, I found that the quality of the code that it was giving me

73
00:08:04,520 --> 00:08:10,520
did result in more errors than I was getting from chat GPT when I did the same project previously.

74
00:08:11,560 --> 00:08:19,640
And the way to fix the errors or its ability to fix its own errors wasn't as good, wasn't close

75
00:08:19,640 --> 00:08:27,320
to being as good. I got to a certain point with this one particular project where I asked it to

76
00:08:28,440 --> 00:08:34,120
fix this code again and again and again, couldn't progress. So I took all of the code it had given

77
00:08:34,120 --> 00:08:40,520
me, put it into chat GPT with GPT-4 and said, fix this code. It fixed it. And then I brought it back

78
00:08:40,520 --> 00:08:47,880
into Gemini Ultra and carried on the project from there. I love it. Gemini Ultra, I love it.

79
00:08:47,880 --> 00:08:54,600
I love it. Gemini Ultra is chat GPT's intern. I don't think that was Google's intention.

80
00:08:58,280 --> 00:09:02,520
And that's kind of indicative of the experience with it. There's just things where it's

81
00:09:05,560 --> 00:09:13,640
quite as good. It's very, very, very good. This is not as slight on it as such. I think Google

82
00:09:13,640 --> 00:09:20,280
have done a tremendous job, but it's still not quite as good as GPT-4. That's the reality of it.

83
00:09:20,920 --> 00:09:27,080
For in fairness, when all the launch videos came out, although given that some of them have been

84
00:09:27,080 --> 00:09:33,880
debunked, how much we can pour credence in them. But they did mention having a specific

85
00:09:34,680 --> 00:09:39,720
model for coding. And that wasn't clear to me at the time whether it was going to be baked into

86
00:09:39,720 --> 00:09:45,960
Gemini Ultra. And in the launch blog, Sundar Pichai does speak about that there are other

87
00:09:47,080 --> 00:09:51,560
models being trained and other product releases coming. So I wonder if there'll be like a dedicated

88
00:09:51,560 --> 00:09:56,680
coding assistant that's maybe better than Gemini Ultra. But given the fact that it's supposed to

89
00:09:56,680 --> 00:10:03,000
be multimodal from the ground up, unlike chat GPT, it's surprising that chat GPT would therefore

90
00:10:03,000 --> 00:10:09,160
be better at coding than a multimodal from the ground up model. But I guess code images, video

91
00:10:09,160 --> 00:10:14,760
text, you know, maybe code is a bit of a unique thing on its own and we'll have to wait and see

92
00:10:14,760 --> 00:10:19,560
what this other model is if indeed it turns out to exist in the way that it sounds like it did.

93
00:10:19,560 --> 00:10:27,480
There are other quirks with it as well. So I had an audio recording, it was just a voice memo

94
00:10:27,480 --> 00:10:35,080
that I'd saved. And I put it into Gemini and asked it to turn that into a blog outline.

95
00:10:35,080 --> 00:10:39,080
And I thought I'll use that blog outline to create a blog. And it was just from a 10 minute

96
00:10:40,200 --> 00:10:48,840
rambling about Zapier and chat GPT. And it failed catastrophically because it said,

97
00:10:49,400 --> 00:10:55,480
I'm sorry, I can't generate images at this moment. And it didn't make any sense to me,

98
00:10:55,480 --> 00:11:00,280
right, because I wasn't asking it nowhere in this text. In fact, I put in a solid

99
00:11:00,280 --> 00:11:07,480
block of text, 10 minutes worth of spoken word. At the end of the prompt, I said, use this spoken

100
00:11:07,480 --> 00:11:13,000
word transcription to write a blog outline detailing this, this, this and this. And

101
00:11:13,000 --> 00:11:18,840
repeatedly, it said it cannot generate images at this moment in time. And it apologized. And

102
00:11:19,720 --> 00:11:24,040
if you ever used Bard, you'll see that it came up with a little drop down to show you alternative

103
00:11:24,040 --> 00:11:31,000
drafts that it also created. Every one of them said the same thing. So somewhere, and I've,

104
00:11:31,640 --> 00:11:41,240
I've reran this in a new chat. And there was one tiny section in this transcript where I mention

105
00:11:41,240 --> 00:11:49,000
using Dali as part of a Zapier workflow to generate an image. But that is like one line

106
00:11:49,000 --> 00:11:52,760
in the middle of this big chunk where at the end of the prompt, I'm very explicit,

107
00:11:52,760 --> 00:11:58,280
I'm very explicit giving it the instruction that I want it to do. And repeatedly, it says,

108
00:11:58,280 --> 00:12:03,160
I can't do that at the moment. And I just had this a couple of times where it was saying things

109
00:12:03,160 --> 00:12:11,960
that it couldn't do or would tell me things that, that weren't quite right. So yeah, that was,

110
00:12:11,960 --> 00:12:17,240
that was frustrating. That aside, there's a lot to like about it. It's super quick. I feel compared

111
00:12:17,240 --> 00:12:25,720
to GPT-4, it generates responses quicker. Admittedly, we have GPT-4 turbo, which is,

112
00:12:25,720 --> 00:12:31,640
you know, pretty quick now. This does still feel quicker. Yeah. But would you not say,

113
00:12:31,640 --> 00:12:35,960
I don't feel like the speed of output of GPT-4 is like a major limiting thing where I'm like,

114
00:12:35,960 --> 00:12:41,320
I'm not going to work with it. It takes, you know, half a second, a second longer to produce an

115
00:12:41,320 --> 00:12:47,480
output. I think it is a little bit annoying. And I do like the idea at some point of just like

116
00:12:47,480 --> 00:12:51,560
getting an instant output, but the moment outputs are about the right speed that I can read at.

117
00:12:51,560 --> 00:12:57,480
So it works quite well for me. But yeah, so I appreciate the speed as being a little bit

118
00:12:57,480 --> 00:13:02,600
of a benefit, but I definitely take quality of output over a slight improvement of speed, I think.

119
00:13:02,600 --> 00:13:08,360
In terms of creative writing or editing content for like blogs and things like that,

120
00:13:08,360 --> 00:13:17,560
I like it. Actually, it has less of those, I think of it as the, well, with chat GPT,

121
00:13:17,560 --> 00:13:21,560
if you ask it to write a blog, it just seems to say we're going to delve, we're going to dive,

122
00:13:21,560 --> 00:13:29,000
we're going to, it uses these phrases that are, just see them so much more with chat GPT than I

123
00:13:29,000 --> 00:13:35,160
see them in real life. Well, certainly more than I would ever write them. And I think there's been a

124
00:13:35,160 --> 00:13:41,720
proliferation of them across the web. This doesn't do that as much. And it gives nice reasons for why

125
00:13:41,720 --> 00:13:45,640
it's phrased something in a particular way, if you ask it to kind of explain why did you describe

126
00:13:45,640 --> 00:13:50,680
it in that way. So I do think that the content writing is pretty good. I thought the marketing

127
00:13:50,680 --> 00:13:55,320
content that I got from it when I was testing it with email campaigns and subject lines and

128
00:13:55,320 --> 00:14:04,120
things like that, all very good. I asked it about how to approach a particular project,

129
00:14:04,120 --> 00:14:08,680
gave it an outline of what I wanted to achieve and the steps involved. And it did a really good job

130
00:14:09,800 --> 00:14:19,480
with giving the step-by-steps as good or even better maybe than the chat GPT. So I thought that

131
00:14:19,480 --> 00:14:29,960
was worth noting. It's context windows 32K, which is the same as GPT-4 and chat GPT. So that's no

132
00:14:29,960 --> 00:14:36,360
real difference there. It can't do image generation at the moment, as it told me repeatedly, even

133
00:14:36,360 --> 00:14:42,200
though I didn't ask it to create an image. I think that will, I would imagine that that will be

134
00:14:42,200 --> 00:14:49,240
coming down the line fairly soon. Google has been making some noise in the image generation arena

135
00:14:49,240 --> 00:14:53,960
over the past few weeks with some new models that it's created. So that's definitely something that

136
00:14:53,960 --> 00:15:01,240
I think that might be being rolled out. I remember seeing a few posts online about Google's image

137
00:15:01,240 --> 00:15:05,640
generation capability starting to make it into these models and that the images were good,

138
00:15:05,640 --> 00:15:11,320
photorealistic, probably better than DORLY 3 if I remember reading. So I think that will come soon.

139
00:15:11,320 --> 00:15:16,440
It's interesting, your take on creative writing. So Ethan Molluck, who does a lot of work in this

140
00:15:16,440 --> 00:15:21,880
area, wrote a blog post about Gemini Ultra having had it for about a month before the rest of us

141
00:15:21,880 --> 00:15:29,080
got it. And I think here's takeaways with similar to yours, right? Poor at coding, probably on par

142
00:15:29,080 --> 00:15:34,440
with text generation for things like creative writing. You did mention that it had a bit of a

143
00:15:34,440 --> 00:15:39,560
more friendly persona and a bit more inclined towards things like conversation and wordplay,

144
00:15:39,560 --> 00:15:45,080
which I think is interesting. Yeah. One of the things I noticed about it very quickly was

145
00:15:45,080 --> 00:15:53,480
its responses are super friendly, actually, and almost like a mentor. So the way that it's,

146
00:15:53,480 --> 00:15:58,360
there was a particular phrase that is like, I'm really excited to get started on this project with

147
00:15:58,360 --> 00:16:03,400
you. When I gave it a brief and asked it to tell me step by step how it would approach it,

148
00:16:03,400 --> 00:16:07,720
it did that. And at the end it was really enthusiastic and said, I'm really excited to

149
00:16:07,720 --> 00:16:12,600
get started on this project with you. But then as I reached a certain point, and it was the coding

150
00:16:12,600 --> 00:16:18,920
one, it was the little app development piece that I wanted it to do, I got frustrated and I said,

151
00:16:18,920 --> 00:16:25,320
this isn't working. I think this project's going to be a failure. And it responded with a piece of

152
00:16:25,320 --> 00:16:31,240
like coaching and it was saying, I'm sorry you feel this way about the project. Let's break down

153
00:16:31,240 --> 00:16:35,240
some of the reasons why you think you might be having these feelings about the project and

154
00:16:35,240 --> 00:16:42,840
how we might be able to overcome it in the next steps. It felt like a real personal dialogue and

155
00:16:42,840 --> 00:16:50,520
a bit of mentorship kind of jeeing me up and telling me that all isn't lost.

156
00:16:50,520 --> 00:16:56,840
So I guess the million dollar question is, are you going to keep your subscription and run it at the

157
00:16:56,840 --> 00:17:05,400
same time as having chatGPT teams or is GPT-4 does everything that Gemini can do and better? So

158
00:17:05,400 --> 00:17:09,080
actually if you had to choose, you'd just choose GPT-4. Like where do you sit at the moment based

159
00:17:09,080 --> 00:17:17,880
on what you've been playing with? If this wasn't my job and I was only going to pick one model,

160
00:17:17,880 --> 00:17:31,880
I would pick GPT-4 or chatGPT above all others at the moment. However, there is some caveats to that.

161
00:17:31,880 --> 00:17:40,120
And that's if you're someone who is personally in the Google ecosystem already, you've got Google

162
00:17:40,120 --> 00:17:49,640
photos, you have got a Gmail account, you use Google Docs and Google Sheets to manage your kind

163
00:17:49,640 --> 00:17:55,480
of personal, you know, maybe you've got a Chromebook and you're kind of all in on the Google

164
00:17:55,480 --> 00:18:04,280
ecosystem. I think for 19 pounds a month, including all of those add-ons that you get, I think I'd be

165
00:18:04,280 --> 00:18:14,760
tempted to go with that. Really? Because you get access to, if GPT-4 is really what, there's some

166
00:18:14,760 --> 00:18:22,200
reason that you really need GPT-4, you can access that via Bing chat. You don't get advanced data

167
00:18:22,200 --> 00:18:31,080
analysis and code interpreter. So if you need those specific elements, get chatGPT, you need that,

168
00:18:31,080 --> 00:18:35,720
right? That's a specific function of Vita that you need. But if you're just looking at the quality

169
00:18:35,720 --> 00:18:44,600
of the output and you're looking at it from a value for money lens, I think Gemini Ultra for 19,

170
00:18:44,600 --> 00:18:49,880
so it's cheaper, 19 pounds a month, plus all of the add-ons. And if you think that you might already

171
00:18:49,880 --> 00:18:56,680
be paying, I think I was paying seven pounds a month for my Google One account. So effectively,

172
00:18:56,680 --> 00:19:04,280
it's 11 pounds or 10 pounds more per month. I think that's a good deal.

173
00:19:04,280 --> 00:19:08,840
You said like a man who loves a subscription, Thierry Martin. I don't disagree with you. I'd

174
00:19:08,840 --> 00:19:15,960
just like to tease you about all your subscriptions and your app sumo addiction. I think I'd be most

175
00:19:15,960 --> 00:19:21,800
tempted to play with it when it comes to workspace. We run on Google Apps, a biostrata in terms of

176
00:19:21,800 --> 00:19:27,640
that's what we use for email. And so the ability to summarise email threads, maybe suggest replies,

177
00:19:27,640 --> 00:19:32,520
I think that's going to be interesting and a potential time saver. If you run on Microsoft,

178
00:19:33,080 --> 00:19:36,840
you're going to use Copilot, probably going to get similar quality outputs by the sounds, I think.

179
00:19:36,840 --> 00:19:41,480
So it might just depend on what ecosystem you're on. It makes me feel better to know that

180
00:19:41,480 --> 00:19:45,880
Ultra is getting up there with GPT-4 because I haven't been on Google Apps this whole time.

181
00:19:45,880 --> 00:19:49,480
There was a bit of a concern that if Google is slow to the party, do we have to migrate

182
00:19:49,480 --> 00:19:56,760
to Microsoft just to get access to all these tools? Paul Roitzer was on the LinkedIn's doing

183
00:19:56,760 --> 00:20:02,200
a breakdown of the value of these tools versus the time that they save you. And, you know,

184
00:20:02,200 --> 00:20:06,520
you look at the time they save you, I would argue that there's also value in the effort of

185
00:20:06,520 --> 00:20:10,200
switching platform if you need to, if the other platforms are not good enough. So that was

186
00:20:10,200 --> 00:20:15,000
definitely on my mind. I think it's important for the listeners to recognise a lot of the things

187
00:20:15,000 --> 00:20:19,720
that plague other models, plague this model. So things like hallucination are still an issue

188
00:20:19,720 --> 00:20:26,040
when you're using Gemini Ultra. One thing I do like about the interface, though, is when it gives

189
00:20:26,040 --> 00:20:29,960
you an output, there's like a little button that you can click. I think it's like a G for Google,

190
00:20:29,960 --> 00:20:35,320
and it will actually go Google parts of its own answer to figure out if it can find information

191
00:20:35,320 --> 00:20:39,880
that either backs up or goes against what it's just told you. So it's like a, it's like

192
00:20:39,880 --> 00:20:45,800
hallucination checking light. For some of the outputs I get, I get there's maybe a thousand

193
00:20:45,800 --> 00:20:50,600
words and 150 of them are highlighted in the green that says, hey, this we can corroborate this with

194
00:20:50,600 --> 00:20:54,680
things we found online. So I don't know what that means for the other 850 words. Right? Is it made

195
00:20:54,680 --> 00:21:00,280
up? Don't know. So it doesn't always work that well. But I've also had outputs where 70% of it

196
00:21:00,280 --> 00:21:04,200
was I think it's green, it might be yellow, the colour that says A is okay, I found some of this

197
00:21:04,200 --> 00:21:09,480
stuff on the web, and occasional red, which is a little bit of a red, which is a little bit of a

198
00:21:09,480 --> 00:21:14,200
red, which means this actually goes against what I found on the web. And I found that button to

199
00:21:14,200 --> 00:21:20,280
actually be quite helpful. Ultimately, that's not better for me than perplexity. Now you've got that

200
00:21:20,280 --> 00:21:24,040
me onto that. So I just using perplexity anyway. So it's more of an oddity for me. I just like,

201
00:21:24,040 --> 00:21:27,640
oh, that's interesting. An AI that checks its own work. That's what we need.

202
00:21:28,360 --> 00:21:34,200
It does have another interesting little UI element as well, which is I've only seen this on the

203
00:21:34,200 --> 00:21:41,480
mobile version though, not sure why. I haven't managed to find it elsewhere. But you can click

204
00:21:41,480 --> 00:21:46,120
on a button and it will ask, you can get it to rewrite the response that it's giving you. So

205
00:21:46,120 --> 00:21:51,560
make it more formal, make it longer, make it shorter, make it more professional. And that's

206
00:21:51,560 --> 00:21:55,240
just right there in the interface rather than having to prompt it to say make that sound more

207
00:21:55,240 --> 00:21:59,880
professional. Yeah, that's kind of similar to what you can get when you're using Bing. I think we're

208
00:21:59,880 --> 00:22:07,960
going to see that in Gmail's AI add-on and Outlook's AI add-on. I think that's going to be. But when I

209
00:22:07,960 --> 00:22:14,840
use those tools, I don't know about you, but it doesn't half love the edge of the continuum.

210
00:22:14,840 --> 00:22:20,520
It's like, make this more chatty. Yo, what up dudes? I've been just chilling. It's like, oh,

211
00:22:20,520 --> 00:22:27,240
crumbs. I didn't mean that chatty. Just like a little bit less stoic prison guard barking that

212
00:22:27,240 --> 00:22:31,560
you had previously just maybe soften it up a bit. So those buttons have seemed to be a brilliant

213
00:22:31,560 --> 00:22:37,080
way of like racing to the extreme without ever actually getting me what I was after in the first

214
00:22:37,080 --> 00:22:43,640
place. So I actually think when these tools learn our styles, personal styles, and I'm sure they will,

215
00:22:44,200 --> 00:22:49,240
that's when it will really help. They be more witty, be more conversational buttons,

216
00:22:49,240 --> 00:22:54,200
and mostly just and fun oddity for me at the moment. They don't actually give me anything usable.

217
00:22:54,200 --> 00:23:02,440
Will Barron Well, thank you for taking us through Gemini Ultra. I think we've given it 21 minutes

218
00:23:02,440 --> 00:23:07,720
or so. That's more than enough for Google to dominate the start of this podcast. Let's move

219
00:23:07,720 --> 00:23:17,400
on to our next story, Martin. You've been looking at some recently published AI readiness index from

220
00:23:17,400 --> 00:23:20,120
Cisco. What's this all about? Martin McGrath

221
00:23:20,120 --> 00:23:25,960
Yeah, they've published a report looking at how corporations around the world are preparing to

222
00:23:25,960 --> 00:23:34,680
integrate and deploy AI within and across their various operations. And there's some stats in

223
00:23:34,680 --> 00:23:42,280
there that won't surprise you. For instance, 84% of companies believe that AI is going to

224
00:23:42,280 --> 00:23:49,640
significantly impact their business. That's not a great surprise at all. Some of the areas

225
00:23:49,640 --> 00:23:56,120
and the recommendations from the report, the areas of insight and the recommendations, should I say,

226
00:23:56,840 --> 00:24:03,720
are worth highlighting. So before we get into that, just a few of the key statistics. 97%

227
00:24:04,280 --> 00:24:11,320
of respondents have seen increased urgency to deploy AI technologies, which is not a massive

228
00:24:11,320 --> 00:24:17,880
surprise when every company that comes out on an earnings call, when they mention AI, their stock

229
00:24:17,880 --> 00:24:29,080
seems to grow a few dollars and pumps the price a little bit. Only 14% feel fully prepared for

230
00:24:29,640 --> 00:24:37,560
AI integration though. And 76% of organizations say that they will require more GPU resources

231
00:24:38,200 --> 00:24:44,440
for AI workloads. So thinking about the actual infrastructure, they're going to have to be

232
00:24:44,440 --> 00:24:52,520
investing. Good news for Nvidia, I guess there. 37% have identified skills gaps in understanding

233
00:24:53,720 --> 00:25:01,320
AI tools and getting their workforce AI proficiency raised across, that is a big

234
00:25:01,320 --> 00:25:07,240
priority for them. And unsurprisingly, generative AI is a focus. 40% of companies are actively

235
00:25:07,240 --> 00:25:13,080
deploying or planning to deploy it. So they're the kind of headline figures for what came out

236
00:25:13,080 --> 00:25:21,720
of the report in terms of takeaways for companies. Well, Cisco identified six pillars of the critical

237
00:25:21,720 --> 00:25:28,760
success factors really for being AI ready. And they said, it's strategy, infrastructure, data,

238
00:25:29,320 --> 00:25:37,240
governance, talent, and culture. So these components range from having a very clearly defined AI

239
00:25:37,240 --> 00:25:43,960
strategy and roadmap, and then owning and having access to the necessary infrastructure to support

240
00:25:43,960 --> 00:25:50,840
those AI workloads, hence those CPUs, ensuring that you've got the right quality of data,

241
00:25:50,840 --> 00:25:59,080
because that's going to impact everything that you do with your workflows. Particularly if you're in

242
00:26:00,440 --> 00:26:06,600
healthcare industry or if you're in the finance sector, having the right data and the right

243
00:26:06,600 --> 00:26:14,520
quality of data, so important. And there's a huge part of any AI project, which is basically data

244
00:26:14,520 --> 00:26:21,880
preparation. And that's something that companies need to get right. And then proper governance,

245
00:26:21,880 --> 00:26:26,920
who has responsibility for what, what are the boundaries, making sure that you've got the right

246
00:26:27,560 --> 00:26:35,160
technology stack. All of these things come into the mix. Finding the right talent, which is

247
00:26:35,160 --> 00:26:40,760
increasingly proving difficult. There's so much demand for AI talent. And then fostering a culture

248
00:26:40,760 --> 00:26:46,920
receptive to AI driven change was the big one. And that requires the right leadership and the

249
00:26:46,920 --> 00:26:53,560
right messaging to communicate that through all of the teams. Surprisingly, the report suggests

250
00:26:53,560 --> 00:27:01,800
that AI could be more of an opportunity for workplace growth than a job replacement threat.

251
00:27:01,800 --> 00:27:09,080
With companies planning significant investments in upskilling their employees in AI technologies,

252
00:27:09,560 --> 00:27:20,440
rather than talking about replacing employees with AI technologies. So despite the moderate and high

253
00:27:20,440 --> 00:27:29,240
urgency to embrace AI, the willingness to adopt varies significantly depending on one's level

254
00:27:29,240 --> 00:27:35,400
within an organization. So what it found was that people lower down organizations are a little bit

255
00:27:35,400 --> 00:27:42,360
more reluctant to embrace the change when the pace from senior management is we need to be all AI,

256
00:27:42,360 --> 00:27:50,040
all go, go, go, go, go. And I think that's reflected in a wider cultural piece outside of

257
00:27:50,920 --> 00:27:57,000
the organization individually, but just a kind of, you know, the macro corporate environment.

258
00:27:57,000 --> 00:28:05,160
There was a really interesting interview with the editor in chief of The Economist. And she sat down

259
00:28:05,160 --> 00:28:13,000
at Davos with Sam Altman and Satya Nadella. And they had a conversation about AI, where it is,

260
00:28:13,000 --> 00:28:18,600
where it's going and all of that. And what was very clear was that the theme of the year at

261
00:28:18,600 --> 00:28:29,240
Davos this year was AI, AI, AI. So clearly corporate leaders are just swimming in this pool of AI hype,

262
00:28:29,240 --> 00:28:35,240
but then cascading that down into the business where you've got, you know, the project managers

263
00:28:35,240 --> 00:28:39,080
that are trying to just get on with the day to day and deliver what they've got to do,

264
00:28:39,080 --> 00:28:44,520
whilst at the same time being told they've got to implement AI into their workflows. There's a

265
00:28:44,520 --> 00:28:51,160
there's a mismatch in appetite for AI intervention.

266
00:28:51,160 --> 00:28:56,920
Yeah, it's a funny one. When I reflect on this report and some of the conversations that we have

267
00:28:56,920 --> 00:29:04,040
on the podcast, I actually think there's going to be not only is it not going to be one size fits all

268
00:29:04,040 --> 00:29:10,440
in terms of how different organizations and different sectors can leverage AI to improve

269
00:29:10,440 --> 00:29:16,440
productivity and efficiency. But just starkly, I'm thinking, if you're an enterprise level company,

270
00:29:16,440 --> 00:29:22,040
the way you need to approach and think about AI as a differentiator and an accelerator for you is

271
00:29:22,040 --> 00:29:26,760
going to be very different from an SME. And it's probably going to be different from the

272
00:29:26,760 --> 00:29:30,840
mid market. And the first thing, and I think there's pros and cons, right? Like if you're

273
00:29:30,840 --> 00:29:35,880
an enterprise level, you probably have lots of access to lots of data. And you do need to think

274
00:29:35,880 --> 00:29:43,400
about format and cleanliness of that data to then drive the production of your own fine tuned models

275
00:29:43,400 --> 00:29:48,120
or even your own models based on the data that you have. If you're an SME, you've probably

276
00:29:49,240 --> 00:29:53,880
got some data, but it's probably so poorly structured as to make it maybe borderline useless.

277
00:29:53,880 --> 00:30:01,720
I don't know. I think if the AI learning tools get good enough to be able to just create insight

278
00:30:01,720 --> 00:30:06,040
from unstructured data, we've talked previously about having a company record all their sales

279
00:30:06,040 --> 00:30:10,120
calls and all their customer service calls. Even if you're a company of 20 people, you could do

280
00:30:10,120 --> 00:30:15,080
that right. But that's unstructured data. So as long as it can start to use that type of unstructured

281
00:30:15,080 --> 00:30:21,800
data to then train models on, I think that's fine. On the smaller side, the change management's

282
00:30:21,800 --> 00:30:24,840
easier because you've got fewer employees that you have to think about how you're going to bring

283
00:30:24,840 --> 00:30:31,160
them together to adopt these tools. Training is probably easier than it is to roll this out across

284
00:30:31,160 --> 00:30:37,640
a 10,000 person organization with offices across the world and different cultures and maybe some

285
00:30:37,640 --> 00:30:42,280
of the tools are really powerful in English, but maybe in local language, maybe they don't work as

286
00:30:42,280 --> 00:30:45,480
well and all the other problems that you're probably going to get. So you step back out of

287
00:30:45,480 --> 00:30:49,640
this for a second, you can see that how you're going to deploy AI in your organization is quite

288
00:30:49,640 --> 00:30:54,360
complicated depending on what type of business you are, what services and products you offer,

289
00:30:54,360 --> 00:31:01,240
structure of your team, size of your team, access to data. We can keep mentioning these reports and

290
00:31:01,240 --> 00:31:05,720
I think they're helpful because they give us all a lens to look through, but you're probably better

291
00:31:05,720 --> 00:31:09,640
off really looking inside your own business against those different facets and thinking,

292
00:31:09,640 --> 00:31:16,200
what does this mean for me? Well, now seems like a relevant time to say go and read my paper published

293
00:31:16,200 --> 00:31:24,520
in October 2022, the month before ChatGPT was even launched, which was about how to deploy AI

294
00:31:24,520 --> 00:31:30,440
in your business for any shape or size of organization. Because it touches exactly on

295
00:31:30,440 --> 00:31:34,120
that. It's about how much resource and how much data you've got in the organization, right?

296
00:31:34,120 --> 00:31:38,120
And if you're a small business with not a lot of resource and not a lot of data,

297
00:31:39,000 --> 00:31:45,000
there's guides in there for ways that you can practically use AI within your organization.

298
00:31:45,000 --> 00:31:48,280
There we go. Alley-oop. I'll put them up. You slam them down.

299
00:31:51,880 --> 00:31:57,320
Cool. Good stuff. Thanks for sharing that, mine. Our next story this week is about Microsoft

300
00:31:57,320 --> 00:32:04,600
Copilot and the recent upgrades to some of the design features and AI image generation features.

301
00:32:04,600 --> 00:32:11,400
So basically, a bit of an improvement to Microsoft Copilot in terms of both its mobile and web

302
00:32:11,400 --> 00:32:19,560
versions, which makes it easier for you to work with images that you've generated using the chat

303
00:32:19,560 --> 00:32:29,240
interface. So it's kind of cool, but not as cool as I would like it to be. So what you can do is

304
00:32:29,240 --> 00:32:36,040
when you're generating images powered by Dually 3, you can now instantly move those into an editing

305
00:32:36,040 --> 00:32:41,000
environment where you can crop them, resize them, you can magic resize them, so you can change the

306
00:32:41,000 --> 00:32:47,880
aspect ratio of them and the AI will fill in the gaps to make that landscape image work as a

307
00:32:47,880 --> 00:32:52,440
portrait and what have you. And all of that's pretty cool. You can do some interesting things

308
00:32:52,440 --> 00:32:58,280
like background blurring and some shifts in the style, which I think is interesting. But when I

309
00:32:58,280 --> 00:33:03,160
first read the story, I thought you're going to have more fine control over AI-driven editing

310
00:33:03,160 --> 00:33:08,120
of things in the image, like, oh, replace that hat, that bowler hat with a cowboy hat.

311
00:33:08,120 --> 00:33:14,520
And we're not quite there yet. I have confidence that in 12 months we'll be able to do that. But

312
00:33:14,520 --> 00:33:18,920
if those of you who saw this story out about on the LinkedIn's, the Twitter's or the press and

313
00:33:18,920 --> 00:33:25,320
thought, oh, now it's finally here, I can be a designer or I'm briefing a robot designer with

314
00:33:25,320 --> 00:33:30,120
text and it's doing cool things to images, not quite there. So you can do some interesting things

315
00:33:30,120 --> 00:33:35,080
and it is a much streamlined workflow. But for those of you that thought it was going to be

316
00:33:35,080 --> 00:33:40,680
like having a digital designer that you were brief and make very subtle amends to images were

317
00:33:40,680 --> 00:33:45,080
not quite there yet. But I still think it's pretty cool. And it's definitely worth checking out if

318
00:33:45,080 --> 00:33:51,160
you do this type of image generation. But yes, not magic. Thor's mine. They also announced

319
00:33:52,120 --> 00:34:00,120
some updates to Designer, which is a kind of, well, it's a graphic design tool, elements of Canva

320
00:34:00,120 --> 00:34:07,240
about it, lots of templates, drag and drop, very WYSIWYG, easy to use. You can export your

321
00:34:07,240 --> 00:34:13,400
generated images into Designer and use those as part of creating a social media post or something.

322
00:34:14,440 --> 00:34:20,440
That's all fine. The thing that I want to make people aware of is it's the first commercial

323
00:34:21,720 --> 00:34:28,040
application that I've seen that users, or at least I think it uses, it certainly looks and

324
00:34:28,040 --> 00:34:38,680
feels and behaves like it. The Segment Anything model from Meta. So you can just click on any item

325
00:34:38,680 --> 00:34:43,160
and it will cut it out and it will select it. You can duplicate it. And it's got that same

326
00:34:44,680 --> 00:34:53,400
feel where you hover over an item and it will kind of lasso it or most of it. And then it will

327
00:34:53,400 --> 00:34:59,160
actually get all of it. And it just, it feels exactly like that implementation of it. So when

328
00:34:59,160 --> 00:35:06,200
we saw that model, I think, was it April time when that was open sourced, whether or not it is that

329
00:35:06,200 --> 00:35:11,400
model, I don't know, but it certainly feels like it behaves like that. So that's a nice little AI

330
00:35:11,400 --> 00:35:17,560
feature added into Designer. Yeah. Adobe's got something like that in what was called Firefly,

331
00:35:17,560 --> 00:35:21,320
but all these things are now having their names changed as they just get baked into products. But

332
00:35:21,320 --> 00:35:28,360
a lot of the demo videos from two, three months ago had Adobe where you click on a suitcase,

333
00:35:28,360 --> 00:35:33,000
automatic lasso the suitcase, delete the suitcase, and it fills in the background like it was never

334
00:35:33,000 --> 00:35:37,480
there. And I think we're just starting to see those emerge. Absolutely going to be driven by

335
00:35:37,480 --> 00:35:42,120
the type of Segment Anything models that you describe and the ability to contextually understand

336
00:35:42,120 --> 00:35:48,840
what's in the image and select it for you. So I guess because of that capability, we're probably

337
00:35:48,840 --> 00:35:55,080
not that far away from speaking to the computer or a text based prompt that says select the suitcase

338
00:35:55,080 --> 00:35:59,720
and remove it. Right. Yeah. Because that's just going to be an action, isn't it? It's a series of

339
00:35:59,720 --> 00:36:06,280
actions, which is click on the suitcase, select it, delete it, generative fill it. That will easily

340
00:36:06,280 --> 00:36:13,000
be done through an AI agent before long. Yeah. Apple released something recently as well that

341
00:36:13,000 --> 00:36:19,960
was similar. They were working with UC Santa Barbara researchers on new sort of image editing

342
00:36:19,960 --> 00:36:27,480
AI that they've released as open source, which is interesting. And it's kind of very similar in that

343
00:36:27,480 --> 00:36:32,600
you have the ability to interact with bits of an image in a conversational way to have it change

344
00:36:32,600 --> 00:36:39,480
the brightness and tweak colors and stuff like that. I think what was interesting about this

345
00:36:39,480 --> 00:36:46,120
model is because of the way it's trained, its ability to influence just the things that you

346
00:36:46,120 --> 00:36:51,080
mentioned conversationally and leave the rest of the image alone, I think is a bit of a different

347
00:36:51,080 --> 00:37:02,440
differentiator for it. So it's kind of drifting towards what I hoped Copilos News was. So if

348
00:37:02,440 --> 00:37:06,920
segment anything is anything to go by, we can expect to see this type of capability in a paid

349
00:37:06,920 --> 00:37:12,280
product within the next six to eight months. But just further evidence that we're drifting towards

350
00:37:12,280 --> 00:37:21,400
that. As a marketer, it's interesting because the further democratization of tools to make it

351
00:37:21,400 --> 00:37:26,840
easier for non-technical specialists to make these types of edits is obviously increasing.

352
00:37:26,840 --> 00:37:31,880
You might write a written brief or call a designer and say, Hey, love the brochure cover,

353
00:37:31,880 --> 00:37:36,200
but can we move this thing here? I want to change the color of this. Now, maybe you do that by

354
00:37:36,200 --> 00:37:41,880
speaking to the computer instead. But the one thing I always fall back on when I'm thinking

355
00:37:41,880 --> 00:37:51,640
about this is it doesn't teach the human user what good design looks like. So it's a bit of

356
00:37:51,640 --> 00:37:56,520
a rubbish in, rubbish out potentially in terms of it's great to be able to ask the computer

357
00:37:56,520 --> 00:38:01,320
and an AI to create an image and then edit specific parts of an image. But if you don't know

358
00:38:01,320 --> 00:38:08,280
about font use, balance, image composition, all these other things that talented designers

359
00:38:08,280 --> 00:38:13,560
are really good at, then probably you'll just have a more official workflow for creating crappy

360
00:38:13,560 --> 00:38:18,360
looking images. And then the other thing is the creative brainstorming that comes behind

361
00:38:19,560 --> 00:38:25,160
creative campaign concepting or brand concepting. Again, I do think you can use some of the tools

362
00:38:25,160 --> 00:38:30,440
to get you started on some interesting ideas, but we're at least to me, it feels miles away from

363
00:38:30,440 --> 00:38:37,480
coming up with really polished creative campaign concepts. So it's interesting and exciting.

364
00:38:37,480 --> 00:38:42,680
And I think to be honest, we're probably going to see that these tools make professional designers

365
00:38:42,680 --> 00:38:47,160
more efficient. I'm not so sure that they're going to make it that much easier for non-designers

366
00:38:47,160 --> 00:38:55,480
to become designers. Do you know what I mean? Yeah. If you're still sticking word art on your

367
00:38:55,480 --> 00:39:02,600
posters and putting things out in Comic Sans, this isn't going to help too much. Love a bit

368
00:39:02,600 --> 00:39:09,240
of Comic Sans. It's just such a happy, such a happy font, but maybe not for the professional

369
00:39:09,240 --> 00:39:17,400
branding arena. Right. Let's move on from image generation. Let's turn to some more AI regulation,

370
00:39:17,400 --> 00:39:22,040
Martin. You've been looking at what's been going on in the European Union. Yeah, just a kind of

371
00:39:22,040 --> 00:39:27,320
headline update really. So we've been following this story as the EU's AI Act has been moving

372
00:39:27,320 --> 00:39:35,560
through the parliamentary process. Well, there was an update last week and the fight, so this is the

373
00:39:35,560 --> 00:39:46,200
pre-final text of the EU AI Act has been endorsed by all 27 EU member states. So the next steps are

374
00:39:46,200 --> 00:39:53,560
that the European Parliament, internal market and civil liberties committees will review it

375
00:39:54,360 --> 00:40:00,920
with a plenary vote provisionally scheduled for April this year. It's hoping that it will be

376
00:40:02,360 --> 00:40:09,640
enforceable, assuming it all goes ahead, in 2026. And I know that Germany's approved it,

377
00:40:09,640 --> 00:40:16,360
which made big headlines. They were one of the ones that everyone was looking towards as one of

378
00:40:16,360 --> 00:40:26,600
the big leader companies, companies, countries in the EU. So yeah, it's taking the next step

379
00:40:26,600 --> 00:40:30,440
forward really through that process. I think if you want to have a look at what the text is going

380
00:40:30,440 --> 00:40:40,280
to largely contain, the version that we've got is going to be it with very minor tweaks around the

381
00:40:40,280 --> 00:40:45,800
edges between now and its final publication. I'm glad you mentioned this story because you reminded

382
00:40:45,800 --> 00:40:52,520
me of something related to callback to Gemini Ultra. So bear with me, listener, we're going to

383
00:40:52,520 --> 00:40:58,440
get there. We have to go around the houses to get there. But Gemini Ultra is going to be baked into

384
00:40:58,440 --> 00:41:03,800
Google Assistant. And I won't say it because I'm surrounded by far too many Google

385
00:41:04,680 --> 00:41:08,040
technologies. One of them is going to tear apart. But obviously you can

386
00:41:09,640 --> 00:41:15,800
speak to your tool by mentioning its name and then you'll get the Assistant technology that

387
00:41:15,800 --> 00:41:20,760
at this point hasn't really improved much in the last five to 10 years. But if you're in the US,

388
00:41:21,320 --> 00:41:27,480
you can switch your Assistant to not being the standard Google Assistant. You can switch it to

389
00:41:27,480 --> 00:41:33,960
being Gemini powered, which is interesting, right? We've been talking about where are all the better

390
00:41:33,960 --> 00:41:38,600
LLM driven smart assistants? Where are they? Well, they're here, but they're not here for Martin

391
00:41:38,600 --> 00:41:44,920
and I in the UK and they're not available in the EU either. And whilst we don't have a clear reason

392
00:41:44,920 --> 00:41:52,200
for that, the rumor is that Google has been paying attention to what's going on with this act before

393
00:41:52,200 --> 00:41:59,720
it takes too many big moves that it thinks could contravene aspects of the act in the EU. So whether

394
00:41:59,720 --> 00:42:05,240
that means we have to wait quite a long time for Gemini powered Assistant or whether this is just a

395
00:42:05,240 --> 00:42:10,200
rumor that isn't really real, you did remind me of it, Martin, when you talked about the story,

396
00:42:10,200 --> 00:42:15,640
because I really, really wanted to try out the new Assistant and I just can't even with VPN,

397
00:42:15,640 --> 00:42:26,120
I can't get it to switch. Which is a surprising move from Google's end, given that we are no

398
00:42:26,120 --> 00:42:37,960
longer in the EU and the regulatory environment of the UK is more innovation friendly, is certainly

399
00:42:37,960 --> 00:42:44,680
on the AI front at least what we're led to believe. Yeah. Well, as I said, it's just a rumor,

400
00:42:44,680 --> 00:42:51,160
might not have any foundation whether or not there's belief in circles that the UK will lean

401
00:42:51,160 --> 00:42:56,120
into following whatever the EU puts in place. I mean, it's tempting to think that the world will

402
00:42:56,120 --> 00:43:01,080
lean into whatever the EU puts in place because the EU has done all the hard lifting of putting

403
00:43:01,080 --> 00:43:06,760
this all together. And you could certainly imagine for a lot of countries, it would be easier to fall

404
00:43:06,760 --> 00:43:12,040
in line with this than it would be to come up with your entirely own approach to doing these things.

405
00:43:12,040 --> 00:43:18,520
But interesting. Well, we should keep an eye on that as it moves its way through. But it sounds

406
00:43:18,520 --> 00:43:24,600
like it's going to pass a little changes. Right. Lungs couple of bits this week, they're less

407
00:43:24,600 --> 00:43:30,200
stories and more just a few discussion topics that I wanted to get Martin's thoughts on.

408
00:43:30,760 --> 00:43:37,880
So the first one in we're going to go back to our friend, Ethan Molloch again here. So probably six

409
00:43:37,880 --> 00:43:45,240
months ago, Ethan ran an interesting experiment with maybe more than maybe it was a year ago,

410
00:43:45,240 --> 00:43:52,600
but he ran an interesting experiment with AI tools as they were emerging to basically create

411
00:43:52,600 --> 00:43:57,880
a variety of content. And I think it was come up with some sort of product launch,

412
00:43:57,880 --> 00:44:03,880
maybe even a new website. And he was able to do all the work that he needed to do in about 30 minutes

413
00:44:03,880 --> 00:44:08,760
using AI tools, which would probably have taken human people like a week or two of time to just

414
00:44:08,760 --> 00:44:16,600
put everything together. So he recently decided to run a similar experiment again, and this time

415
00:44:16,600 --> 00:44:24,600
to show what he could do using these tools, but in just a minute or as he says in 59 seconds. So

416
00:44:24,600 --> 00:44:28,680
he undertook a range of tasks all at the same time. And if you can find this online, it's worth

417
00:44:28,680 --> 00:44:33,480
going and watching the screen capture video of his desktop because he's split it into five parts.

418
00:44:33,480 --> 00:44:40,040
So you can see all of these things running at the same time. So firstly, he uses Microsoft Copilot

419
00:44:40,040 --> 00:44:45,240
to turn an AI written Tesla business case file into a PowerPoint presentation. So it's a word

420
00:44:45,240 --> 00:44:51,000
to PowerPoint transition. He then uses Copilot and Microsoft to devise a detailed syllabus for a

421
00:44:51,000 --> 00:44:55,720
six session intro to entrepreneurship class, complete with table summaries, assignments,

422
00:44:55,720 --> 00:45:02,600
and the grading criteria. He used ChatGPT with a custom GPT made called trend analyzer

423
00:45:02,600 --> 00:45:08,680
to identify and showcase some on trend designs for, I think it's for clothing.

424
00:45:09,480 --> 00:45:13,880
He applied ChatGPT with his own product launch GPT to automate the launch process

425
00:45:13,880 --> 00:45:21,880
of a Wharton interactive products at Wharton Business School. And then he used Bing to draft

426
00:45:21,880 --> 00:45:28,360
a market research study on the VR and AI, VR and AR device market in the style of a top consulting

427
00:45:28,360 --> 00:45:32,760
firm. And his biggest challenge in all of this process was that originally Bing said no,

428
00:45:32,760 --> 00:45:36,600
and he managed to convince it to do it. Typical there of large language models.

429
00:45:37,400 --> 00:45:41,240
So find the video online if you can, because he does all these. He clicks in all the different

430
00:45:41,240 --> 00:45:47,560
windows and within a minute he's got outputs from all of them. And as he says in his blog post,

431
00:45:47,560 --> 00:45:50,680
where he summarizes what he did, the quality of the outputs was surprisingly high,

432
00:45:52,440 --> 00:45:57,400
which is even a big improvement on the last experiment he ran, because we've got these

433
00:45:57,400 --> 00:46:02,760
better models now. So it's quite an interesting one. And as much as this is a big leap forward,

434
00:46:03,480 --> 00:46:06,760
one thing he highlights in his blog, and this is what I want to get your take on Martin, is

435
00:46:07,640 --> 00:46:15,400
up until now a lot of knowledge work was about words, writing reports, creating reports,

436
00:46:15,400 --> 00:46:19,800
writing emails, the fact that that takes time and was considered high value work.

437
00:46:20,520 --> 00:46:25,560
If you can now hit a series of buttons and get all of these tasks accomplished to a reasonable

438
00:46:25,560 --> 00:46:30,680
level in 59 seconds, where your job is then to go in and edit and increase the quality of those

439
00:46:30,680 --> 00:46:39,320
things. How does that change work? How do we understand what is good work and what is bad

440
00:46:39,320 --> 00:46:44,920
work when AI is doing so much of the lifting? Where's the human bringing the magic to this?

441
00:46:45,960 --> 00:46:52,120
For lots of jobs that are based on time investment, like agencies, honestly, a lot of agencies are

442
00:46:52,120 --> 00:46:56,680
based on a time model, lawyers, lots of things are based on how much do I charge for this based

443
00:46:56,680 --> 00:47:01,560
on how long I think it would take my team to do. We talk about this a lot, but it was just so

444
00:47:01,560 --> 00:47:06,360
interesting to see all of that work done at a reasonable quality in 59 seconds.

445
00:47:06,360 --> 00:47:08,600
What's your take on some of these questions, Martin?

446
00:47:10,360 --> 00:47:17,320
Well, it makes me think of somebody I was speaking to at the weekend. They worked for a large

447
00:47:17,320 --> 00:47:26,600
corporate, big engineering company and they are not allowed to use LLMs at all. They're all locked

448
00:47:26,600 --> 00:47:33,560
down. And this person I was speaking to, an engineer, had this particular job which required

449
00:47:33,560 --> 00:47:40,760
him to... It was basically, he knew it was going to be his entire afternoon copying and pasting data

450
00:47:40,760 --> 00:47:47,480
from spreadsheet sources, like from here to here, here to here. And that was going to be

451
00:47:48,600 --> 00:47:54,920
an entire afternoon spent doing this part of his task so that he could then run some models and

452
00:47:54,920 --> 00:48:04,520
do some analysis. And he's not allowed to use ChatGPT, but he went on his phone and did it

453
00:48:04,520 --> 00:48:12,440
on the side and said, ChatGPT, give me a VBA script that will enable me to execute this thing.

454
00:48:12,440 --> 00:48:18,200
And he said within 30 seconds, he had the script and he implemented it and it did his four hours

455
00:48:18,200 --> 00:48:27,480
work in about a minute. And the whole process, start to finish, was two or three minutes. At the

456
00:48:27,480 --> 00:48:31,880
time it took him to write the prompt, to copy and paste it, email it to himself. He says, if I was

457
00:48:31,880 --> 00:48:36,680
just... If we just had this as part of our workspace, think about how much time that would have... I

458
00:48:36,680 --> 00:48:41,160
mean, it was already saving a huge amount of time, but if everybody in the organization had this,

459
00:48:41,160 --> 00:48:48,120
rather than being in an organization where they are scared to open this up, for whatever reason,

460
00:48:48,120 --> 00:48:53,480
I think there's data and compliance and IP related concerns, all of the stuff that we've spoken about

461
00:48:53,480 --> 00:48:59,560
in the past. But that just made his job so much easier. And the output that he was going to get

462
00:48:59,560 --> 00:49:06,360
to in his particular instance still required some human intuition, but that data prep piece,

463
00:49:06,360 --> 00:49:15,400
literally copying and pasting, was going to be done for him in seconds. To the wider point about

464
00:49:15,400 --> 00:49:27,720
knowledge and time, we've said this when GPT-4 came out and we saw lawyers, we could see that

465
00:49:27,720 --> 00:49:35,880
GPT-4 could pass the bar exam with the score that is equivalent to top 10% of humans that sit it.

466
00:49:37,080 --> 00:49:44,200
The thing that makes solicitors and lawyers so highly respected is the time, effort, energy

467
00:49:44,200 --> 00:49:49,320
they've put into learning their subject matter. And now everybody has an expert in their pocket.

468
00:49:49,320 --> 00:49:58,280
That can do it. So it's going to diminish some of the value of that work. It has to. There's an app

469
00:49:58,280 --> 00:50:08,680
called Robin.ai. I don't know if you've seen this one. And it's been built using Claude from

470
00:50:08,680 --> 00:50:17,640
Anthropic. And it does contract analysis. And the whole purpose of it is to analyze all of the

471
00:50:17,640 --> 00:50:24,040
clauses within a contract and then help you write the clauses that you want adding into the contract

472
00:50:24,040 --> 00:50:34,120
and help you basically negotiate better using AI. That's something that previously a lawyer would

473
00:50:34,120 --> 00:50:41,640
have charged you, I mean, what do lawyers charge? Hundreds of pounds, thousands of pounds per hour

474
00:50:41,640 --> 00:50:47,080
to sit and go through line by line, checking over every clause. Now we've got AI that can do that,

475
00:50:47,080 --> 00:50:52,840
explain every clause to you and it'll take... I don't have the answer to the question of what does

476
00:50:52,840 --> 00:50:59,640
this mean for work. I just know that it is going to, it's going to change what we perceive as

477
00:51:00,280 --> 00:51:10,680
the value of that kind of work, but it's also going to make as much more productive as possible.

478
00:51:10,680 --> 00:51:17,880
Much more productive. It's going to empower individuals in so many ways that we... It's kind

479
00:51:17,880 --> 00:51:25,560
of hard to imagine really. If everybody has access in their pocket to a legal advisor that is

480
00:51:26,840 --> 00:51:36,600
pretty good, right? Really high level or a doctor or a tutor or a mentor, it's going to change the

481
00:51:36,600 --> 00:51:43,000
way that we think about how we approach and how we value outputs and tasks within the workplace.

482
00:51:44,040 --> 00:51:51,800
Yeah, it's a bit of an Ethan Molyk episode, isn't it? Ethan Molyk and his team released another

483
00:51:51,800 --> 00:51:57,640
paper where they studied consultants at Boston Consulting Group. It's a pretty robust study by

484
00:51:57,640 --> 00:52:02,760
looks of things with large numbers of consultants split into two groups. One group had access to AI

485
00:52:02,760 --> 00:52:09,320
and the other group didn't. One of the sort of findings, it's more of a hypothesis that came

486
00:52:09,320 --> 00:52:17,960
from the analysis, so it was not robustly tested yet, was does access to AI make the smartest,

487
00:52:17,960 --> 00:52:24,840
most capable people better or does it take people who are new in their careers or maybe lower skill

488
00:52:24,840 --> 00:52:31,560
levels and bring them up higher? So does it compress the average? Fundamentally, there's no

489
00:52:31,560 --> 00:52:38,520
reason why both those things can't be true, right? But with AI still prone to hallucination,

490
00:52:39,640 --> 00:52:48,440
my feeling is AI can help do a lot of initial drafting type work. If you're in knowledge work

491
00:52:48,440 --> 00:52:52,200
and a lot of what you're doing is producing insights that end up in written form,

492
00:52:52,200 --> 00:53:01,720
somebody who understands the topic deeply has to review those pieces to look for hallucinations,

493
00:53:01,720 --> 00:53:06,200
but also just to make sure, are we talking about the things that are important for this report?

494
00:53:06,200 --> 00:53:11,000
Right? Because even when you try and give AI a lot of context, you can't give it all the things you

495
00:53:11,000 --> 00:53:16,840
know when you've done something for 20 years, like, and it doesn't have experience like humans have

496
00:53:16,840 --> 00:53:23,880
experience, right? So I'm drifting towards, and this is a deep fear of mine, really, but I'm drifting

497
00:53:23,880 --> 00:53:33,800
towards, could it gobble up intern and graduate roles where people coming into the workforce

498
00:53:33,800 --> 00:53:39,960
from university or by other mechanisms would usually get their training, but now they don't

499
00:53:39,960 --> 00:53:47,160
get it because AI can do some of that entry-level work and at $20 a month, it's cheaper than hiring

500
00:53:47,160 --> 00:53:52,520
a grad, which sounds amazing, like commercially, like, let's say you're the FD, you're like,

501
00:53:52,520 --> 00:53:57,560
well, yeah, okay, this is an obvious choice. But sooner or later, we're going to run out of the

502
00:53:57,560 --> 00:54:02,280
experts to actually validate the outputs from the AI because we won't have trained any juniors

503
00:54:02,280 --> 00:54:08,120
because we would have let AI be the juniors. So we won't have people to validate the outputs anymore,

504
00:54:08,120 --> 00:54:16,520
whilst also creating a bit of a socioeconomic meltdown by not giving kids jobs in graduate

505
00:54:16,520 --> 00:54:22,520
positions. So it's kind of interesting, if I look at the way capitalism is incentivized or

506
00:54:22,520 --> 00:54:30,920
incentivized businesses, it's almost like, don't choose the easy route today because you may pay

507
00:54:30,920 --> 00:54:36,360
for it tomorrow. I guess the only bet would be if AI can accelerate fast enough as to not

508
00:54:36,360 --> 00:54:42,120
make that a problem. But yeah, I think when I look at it, my worry is more for those entry-level

509
00:54:42,920 --> 00:54:47,960
graduates and, you know, I'm in my early 40s, I don't have any kids of myself, but I know a

510
00:54:47,960 --> 00:54:54,680
number of people who've got kids of that 15 to 20 years old and they're asking me what should they

511
00:54:54,680 --> 00:54:59,320
do? And it's, you know, I'd be worried if I had kids of that age, to be honest, because I think

512
00:54:59,320 --> 00:55:03,000
they've had to deal with the pandemic first during their schooling.

513
00:55:03,000 --> 00:55:08,440
Deal with the pandemic first during their schooling. And now potentially, AI could come

514
00:55:08,440 --> 00:55:12,680
and gobble up some of those entry-level jobs that they need to kickstart their careers.

515
00:55:12,680 --> 00:55:21,560
Yes. Equally, because of how good LLMs can be at mentoring and teaching,

516
00:55:22,680 --> 00:55:29,240
I think there's also a case that it can help people become experts quicker as well.

517
00:55:29,240 --> 00:55:35,400
So it could almost help them to become the experts that they need to review the documentation. Like

518
00:55:35,400 --> 00:55:40,920
where I was talking about the Gemini Advanced had this kind of mentorship role where it was telling

519
00:55:40,920 --> 00:55:44,920
me how to approach a project in a slightly different way or how to think about things in a

520
00:55:44,920 --> 00:55:51,480
slightly different way. If we embrace and lean into the idea and the opportunity of

521
00:55:52,200 --> 00:55:58,760
AIs being personal mentors and assistants, I think we could actually start to get to a place

522
00:55:58,760 --> 00:56:08,280
to a place where taking on junior members of staff that are well-versed in how to engage

523
00:56:08,280 --> 00:56:20,600
and interact and use an LLM could be less risky or more, what's the word, deliver a better

524
00:56:20,600 --> 00:56:27,560
return on that hire quicker than it would do otherwise. Because if I was, at the moment,

525
00:56:27,560 --> 00:56:36,280
I don't hire any staff. If I was taking someone on, that would be a big consideration for me.

526
00:56:36,280 --> 00:56:42,600
It's like how comfortable are you using ChatGPT? Because actually, I don't really, as long as you

527
00:56:42,600 --> 00:56:48,600
can figure that out and you're a bit of a tinkerer and you can ask the right questions and you've

528
00:56:48,600 --> 00:56:57,400
got the right mindset, you can GSD, right? Get shit done. And that, I would be more prepared

529
00:56:57,400 --> 00:57:04,920
to take on somebody that was in a grad role with that mindset and with that toolkit than taking on

530
00:57:04,920 --> 00:57:10,280
someone that maybe had a little bit more experience but was a bit reticent and not prepared to embrace

531
00:57:11,320 --> 00:57:16,600
ChatGPT. I think that's the critical piece. Ironically, I think that's always been the

532
00:57:16,600 --> 00:57:23,320
critical piece, honestly. And when I answer that question, I do answer something similar to what

533
00:57:23,320 --> 00:57:27,960
you've said, which is people who are curious problem solvers who make the best use of the

534
00:57:27,960 --> 00:57:32,760
tools around them will always succeed. Especially against the old adage of,

535
00:57:34,280 --> 00:57:40,520
I don't have to outrun the bear, I just have to outrun you. So I do think those people will

536
00:57:40,520 --> 00:57:48,440
still be successful. When I was in the lab, I hate my lab colleagues, don't listen to this podcast.

537
00:57:48,440 --> 00:57:54,680
When I was in the lab, I used to wear a t-shirt that said, just f-ing Google it on it. Because

538
00:57:54,680 --> 00:58:01,240
I used to get so many questions where literally people would watch me Google it and then I would

539
00:58:01,240 --> 00:58:05,880
answer their question and I'm like, dude, just Google it, right? That's what I'm going to do.

540
00:58:05,880 --> 00:58:10,680
You can do it and you don't have to bother me. I didn't have it that often, to be honest. I just

541
00:58:10,680 --> 00:58:17,560
thought the t-shirt was funny. But I think that just ChatGPT is the equivalent, right? So I would

542
00:58:17,560 --> 00:58:25,880
argue we've had the tools for those types of curious self-starting problem solvers to excel.

543
00:58:25,880 --> 00:58:31,640
We've had them for best part 15, 20 years and this is just the latest version of them.

544
00:58:32,760 --> 00:58:37,400
I think they're powerful, but I think they're prone to surfacing the wrong information like

545
00:58:38,840 --> 00:58:42,600
someone using Google who's not sure how to validate the outputs would be.

546
00:58:42,600 --> 00:58:50,200
To add a little extra part to what you said, for me, one of the biggest challenges post-pandemic

547
00:58:50,680 --> 00:58:57,880
is training junior people because they're not learning... I don't know how to describe this

548
00:58:57,880 --> 00:59:05,880
adequately, so I'm just going to do my best. The juniors are learning more slowly and they're not

549
00:59:06,680 --> 00:59:11,400
adopting skills and they don't have the insights that they might have had pre-pandemic

550
00:59:11,400 --> 00:59:16,520
two, three, four, five years into their career journey. If I had to guess why, I would say it's

551
00:59:16,520 --> 00:59:20,920
all those little conversations that you see happening in an office environment that you're

552
00:59:20,920 --> 00:59:26,280
learning from by our osmosis where people with more experience than you are having a conversation

553
00:59:26,280 --> 00:59:29,720
about something that's somewhat related to what you're doing that you're learning from.

554
00:59:29,720 --> 00:59:35,720
I think they're massively missing out on those and I still think LLMs are great at producing

555
00:59:35,720 --> 00:59:41,880
information but I'm not sure they produce insight. They don't have the hard-earned experience of,

556
00:59:42,520 --> 00:59:46,440
yeah, but if you do it in that way in this particular context it's going to go wrong because

557
00:59:46,440 --> 00:59:50,600
of this and how do I know that? Because six years ago it got burned pretty badly when I messed that

558
00:59:50,600 --> 00:59:57,000
up. LLMs don't have any of that so they can't transmit that and that would be my fear. So

559
00:59:57,000 --> 01:00:04,200
I completely agree that a savvy, curious, problem-solving driven grad with an LLM is going

560
01:00:04,200 --> 01:00:12,840
to be very powerful than someone who doesn't have it but I still just worry that that lack of insight

561
01:00:12,840 --> 01:00:19,080
is where the value is going to be lost and how to get that baked back into the system.

562
01:00:19,080 --> 01:00:23,640
Well, yeah, that's for GPT-5 to figure out.

563
01:00:25,640 --> 01:00:32,520
Yeah, I was joking with a mate of mine about how awesome it would be if everyone could

564
01:00:32,520 --> 01:00:37,880
take a year off work, like literally everyone, find all the best people in the world in their

565
01:00:37,880 --> 01:00:45,160
domains and have them do the human reinforcement feedback on a model based on all of the deep

566
01:00:45,160 --> 01:00:52,200
experience that they have. Take the best surgeons, take the best artists, take the best politicians,

567
01:00:52,200 --> 01:00:57,880
take the best lawyers, marketers, business people, entrepreneurs, you name it, and have them do the

568
01:00:57,880 --> 01:01:03,400
reinforcement because then it's almost like humanity spends a year together training AI by

569
01:01:03,400 --> 01:01:09,400
having our very best and brightest be the trainers and the mentors for the AI because I think you'd

570
01:01:09,400 --> 01:01:12,840
end up with something quite cool and interesting after that.

571
01:01:14,200 --> 01:01:20,520
I have heard some AI scientists discussing this exact problem and the problem with reinforcement

572
01:01:20,520 --> 01:01:25,160
learning from human feedback being that you've got to have the right humans giving the feedback

573
01:01:25,160 --> 01:01:32,440
because there's no point having me or you giving feedback on responses about brain surgery.

574
01:01:33,320 --> 01:01:39,160
We are not the right people to be doing that, but equally trying to get brain surgeons sitting

575
01:01:39,160 --> 01:01:49,080
there rating AI responses doesn't feel like a great use of their time or at least feels like it

576
01:01:49,080 --> 01:01:54,120
would be quite an expensive process to go through to get hundreds of people to do the

577
01:01:54,120 --> 01:01:59,480
work to get hundreds of brain surgeons that just short term responses.

578
01:02:03,640 --> 01:02:07,080
So that's why we all need a year off. Who do I speak to about this, Martin?

579
01:02:07,960 --> 01:02:11,160
You must know someone. How do we get a year off for everyone?

580
01:02:11,800 --> 01:02:13,320
That's Sam A on Twitter.

581
01:02:14,680 --> 01:02:17,960
Maybe that's why he's getting the seven million, seven trillion I should say.

582
01:02:17,960 --> 01:02:24,120
Seven trillion, yeah. Currently trying to secure funding for chip manufacture, fusion based energy

583
01:02:24,120 --> 01:02:30,200
systems and all the things that he feels he needs to create Skynet. Maybe we just put all of that

584
01:02:30,200 --> 01:02:34,520
into getting all the experts that we need and paying for them to have six months off work so

585
01:02:34,520 --> 01:02:41,480
that we can train all these AIs to be ninjas. If your investment request doesn't start with a

586
01:02:41,480 --> 01:02:49,480
two, then what's even the point? Let's get out. I have got an idea for a bubble app that I do

587
01:02:49,480 --> 01:02:57,320
think would do really well with a trillion dollars of funding. I'm going to spend a hundred dollars

588
01:02:57,320 --> 01:03:04,440
less than a trillion. All of that's going to be spent on paid advertising and a hundred dollars

589
01:03:04,440 --> 01:03:10,680
that we spent on trying to create the tool. Anyway, let's respect our audience's time,

590
01:03:10,680 --> 01:03:14,600
Mike. We're going to move on. Last segment here is going to be, we're bringing back tool of the week.

591
01:03:14,600 --> 01:03:18,200
Are we bringing it back forever? Probably not. Are we bringing it back today? Yeah,

592
01:03:18,200 --> 01:03:21,960
because Martin's been playing with Claude for Sheets and he loves it and he wants to take us

593
01:03:21,960 --> 01:03:27,880
through it. Well, that about sums it up to be honest, but to put more flesh on the bones,

594
01:03:27,880 --> 01:03:35,080
I have been playing with Claude for Sheets and this is a login or should I say add-on,

595
01:03:35,080 --> 01:03:41,240
it's a Google Sheets add-on developed by Anthropic themselves. It's a first party

596
01:03:41,240 --> 01:03:49,560
plugin which unlike ChatGPT and GPT for Sheets, well, there isn't one for that. There is a GPT

597
01:03:49,560 --> 01:03:54,520
for Sheets, but it's third party plugin and you have to pay for the privilege now of using that.

598
01:03:55,480 --> 01:04:00,120
It was free for a year and that's actually what made me check this one out because

599
01:04:00,120 --> 01:04:05,400
it's been available for a while. It's well documented on the Anthropic website and I thought

600
01:04:05,400 --> 01:04:10,680
I'd give it a go. It's dead easy to set up. All you need is your API key and you can then connect

601
01:04:10,680 --> 01:04:18,600
it to all of your Sheets. When you use it, you just input equals Claude, open brackets,

602
01:04:18,600 --> 01:04:24,120
and then you write your prompt and use your various parameters. But as it's a spreadsheet,

603
01:04:24,120 --> 01:04:35,560
you can nest the prompt and the function inside other formulas. So you can have if this, then that.

604
01:04:35,560 --> 01:04:41,720
You can have if statements and run a prompt according to different criteria or different

605
01:04:41,720 --> 01:04:49,640
conditions of your spreadsheet. There's actually three types of input. So you can have Claude,

606
01:04:49,640 --> 01:04:55,880
which is basically the user sends a message and gets a response. You've got one called Claude Free

607
01:04:56,680 --> 01:05:03,400
and with Claude Free, what you're doing is you're having it more as a conversation starter. So you

608
01:05:03,400 --> 01:05:09,640
can actually have more control over the response that you get back from Claude. So where Claude

609
01:05:09,640 --> 01:05:15,000
might have a bit of a habit of starting its response with a bit of preamble, something like

610
01:05:15,000 --> 01:05:19,880
here is the summarization of the text that you have requested. You don't want that. So you can

611
01:05:19,880 --> 01:05:28,280
put that at the start of the response from it so it doesn't give that. Or you can make it say

612
01:05:28,280 --> 01:05:33,160
anything that it wants at the start and then the cell populates with the rest of the response from

613
01:05:33,160 --> 01:05:41,160
the assistant. And then there's a third setting, which I can't remember what it's called, seems

614
01:05:41,160 --> 01:05:48,520
to do a similar thing. It's more for message based back and forth. You have full control over

615
01:05:49,640 --> 01:05:59,480
the API settings. If you've used chat GPT, you'll know that you only have the input of the text.

616
01:05:59,480 --> 01:06:06,280
You can't control settings like the temperature, which controls how creative or constrained chat

617
01:06:06,280 --> 01:06:14,600
GPT is with its response. You can't control the token count except by saying, do it in less than

618
01:06:14,600 --> 01:06:19,400
400 words or something like that. You can put it into the prompt, but with Claude for Sheets,

619
01:06:19,400 --> 01:06:25,640
you can control the token length. You can control the temperature to make it more creative or more

620
01:06:25,640 --> 01:06:33,800
precise. And you can also control, well, we'll see the other setting. There's another setting as well,

621
01:06:33,800 --> 01:06:39,320
but it escapes me now. But it's super easy to use. If you're familiar with spreadsheets

622
01:06:40,040 --> 01:06:50,760
and you're quite comfortable using Excel or Google Sheets, check it out. Within 15 minutes

623
01:06:51,560 --> 01:06:58,440
of playing around with it, you'll see huge value, particularly if you work in SEO,

624
01:06:58,440 --> 01:07:05,640
if you work in e-commerce. In fact, I had been playing with this and then did a workshop last

625
01:07:05,640 --> 01:07:11,320
week getting started with chat GPT. At the end of the session, I spoke about integrations and tools,

626
01:07:11,320 --> 01:07:16,040
and I mentioned this particular plugin. Yes, I know it's not chat GPT as Claude, but

627
01:07:16,600 --> 01:07:19,720
people were interested. And one of the people that was on this session,

628
01:07:21,080 --> 01:07:24,840
he's got an e-commerce store, hundreds of products, and he's

629
01:07:24,840 --> 01:07:29,880
moving from one platform to another. So he's changing his content management system,

630
01:07:30,840 --> 01:07:36,360
e-commerce platform. As such, he needs to reformat all of his product descriptions

631
01:07:36,360 --> 01:07:42,840
to match the new template that he's using. And he said it's hundreds of them and he's

632
01:07:42,840 --> 01:07:48,120
kind of dreading it. I said, well, have you got them all on a spreadsheet? He said, yeah,

633
01:07:48,120 --> 01:07:52,840
I've exported all of that, but now I've got to sit down and do the rest of the work. And he said,

634
01:07:52,840 --> 01:07:59,400
all of that, but now I've got to sit down and just do it all. And I showed him the plugin and how it

635
01:07:59,400 --> 01:08:10,200
could work. And we estimate that with, well, maybe like 10, 15 minutes finding the right prompt and

636
01:08:10,200 --> 01:08:15,480
playing around with the right prompt. But once he's got that, completing the task less than a

637
01:08:15,480 --> 01:08:23,480
minute, and then he just needs to upload the file to his products. A massive game changer. Yeah.

638
01:08:23,480 --> 01:08:29,480
And we've talked about some of these plugins before. I mean, one of the pains of the GPT-4

639
01:08:29,480 --> 01:08:34,360
based plugin is it hangs all the time, right? Because I think it's a 30 second timeout. So

640
01:08:34,360 --> 01:08:38,760
we always got fed up of using it because it hangs. What's the clawed timeout like?

641
01:08:38,760 --> 01:08:44,600
So the limitation, that 30 second timeout that you're referring to is a Google Sheets

642
01:08:45,720 --> 01:08:50,280
restriction. So anytime it makes an external call to an external API,

643
01:08:51,160 --> 01:08:54,040
if it doesn't get the response back in 30 seconds, it just errors.

644
01:08:56,600 --> 01:09:05,480
So clawed has a setting within it, or clawed for Sheets has a setting within it, to rerun any errors.

645
01:09:05,480 --> 01:09:11,880
And it will just rerun those cells. It also has a state called deferred. So sometimes,

646
01:09:11,880 --> 01:09:16,840
let's say you've got 100 cells that want to be updated, because of the API limit,

647
01:09:17,400 --> 01:09:22,440
it will restrict that and it will put some of them in a queue or it will put some of them as deferred.

648
01:09:22,440 --> 01:09:30,440
So you can rerun it there. I've, having played extensively with GPT-4 through the GPT for Sheets

649
01:09:30,440 --> 01:09:40,280
plugin and the clawed for Sheets plugin, I can say that the response time from clawed is far and away

650
01:09:40,280 --> 01:09:46,280
quicker than GPT-4. And that's the other parameter that you can change in the formula. You can change

651
01:09:46,280 --> 01:09:55,080
the model, of course. You can choose clawed 2.1, which is their most powerful and capable model,

652
01:09:55,080 --> 01:10:02,920
or you can choose the quicker clawed 1 turbo, which is very quick, but obviously not as powerful

653
01:10:02,920 --> 01:10:13,560
and as capable as clawed 2.1. Even clawed 2.1, where I've had some quite long inputs with quite

654
01:10:13,560 --> 01:10:21,320
detailed outputs, I've still found that the error messages are not an issue. Whereas I've got a

655
01:10:21,320 --> 01:10:27,400
project that I'm actively working on at the moment and I've had to switch off GPT-4 with GPT for

656
01:10:27,400 --> 01:10:34,040
Sheets because it errors on at least 20% of them and it's just not, you know, not, you can't do it,

657
01:10:34,040 --> 01:10:38,120
right? Yeah, it's not usable. Yeah, that was the reason for the question, because that was a pain

658
01:10:38,120 --> 01:10:45,240
and we've talked about use cases of this in terms of setting up customized prospecting emails at

659
01:10:45,240 --> 01:10:49,800
scale for hundreds, if not thousands of prospects where you're able to find a way to get ChatGPT

660
01:10:49,800 --> 01:10:56,920
to customize writing the outbound email and a bunch of other stuff. And I think as clawed

661
01:10:56,920 --> 01:11:02,920
2.1's writing output is pretty high quality, maybe not quite GPT-4, but not far off, I can see why

662
01:11:02,920 --> 01:11:07,800
that would be attractive. And if it doesn't hang all the time, which basically was making

663
01:11:07,800 --> 01:11:13,720
the GPT plugin for Sheets unusable, let's be honest, that sounds like a big step. Are you

664
01:11:13,720 --> 01:11:17,720
paying for the API to access this and how much does it cost to do these types of things?

665
01:11:17,720 --> 01:11:22,840
So at the moment, you only start paying for API usage when you upgrade to a production ready.

666
01:11:22,840 --> 01:11:32,360
So anybody can sign up to Clawed API with Anthropic and get a free account and you get 2000,

667
01:11:32,360 --> 01:11:40,840
was it? No, was it? Two million or two and a half million tokens per day or something before you

668
01:11:40,840 --> 01:11:47,640
have to start even looking at a free version. So if you're using it for yourself, you're paying

669
01:11:47,640 --> 01:11:53,480
for it. And if you're using it for yourself in these types of experiments or even for internal

670
01:11:53,480 --> 01:11:58,600
processes, you're probably just going to be able to operate for free. Correct. Yeah, certainly for

671
01:11:58,600 --> 01:12:04,440
the time being. Interesting. It's time for me to go installing this again and see if I can get some

672
01:12:04,440 --> 01:12:08,280
of the cool stuff I've been trying to get done over the last 12 months for some of these plugins

673
01:12:08,280 --> 01:12:12,600
working because I gave up because they were hanging all the time. But if I was, I would say

674
01:12:12,600 --> 01:12:18,680
there's actually a guide from Clawed and Anthropic about prompt engineering specifically for Sheets

675
01:12:18,680 --> 01:12:24,200
and they've got a whole Google Sheets workbook with exercises that you can work through and do it.

676
01:12:24,200 --> 01:12:33,880
But just on the quality output piece, if you are looking for those GPT-4 level outputs, assuming

677
01:12:33,880 --> 01:12:39,160
that the output you're after isn't a thousand words long, if it's, you know, 100 words or so,

678
01:12:39,160 --> 01:12:48,120
maybe it's a summary of text, take a GPT-4 summary or take a few summaries and give that as a,

679
01:12:48,120 --> 01:12:57,240
give that as part of the prompt. And Clawed too is then getting few shot prompt trained. Right.

680
01:12:57,240 --> 01:13:05,960
On GPT-4 quality outputs. And it does a very good job of matching that thereafter. So if you add it

681
01:13:05,960 --> 01:13:11,000
in your prompt, you're basically saying this is what good looks like. Apply this to your outputs.

682
01:13:12,200 --> 01:13:17,720
Interesting. Well, there we have it. Tool of the week made a comeback. If you are trying to figure

683
01:13:17,720 --> 01:13:24,040
out some next level automated uses of chat GPT and Clawed and other large language models until

684
01:13:24,040 --> 01:13:29,720
agents make their appearance, hey, next week, next month, who knows, then these things are good to

685
01:13:29,720 --> 01:13:34,120
play with. Martin, it's always a pleasure hanging out with you. Thanks for your time today. Good

686
01:13:34,120 --> 01:13:40,280
conversation. Yeah. Looking forward to the next. Take it easy, buddy. Speak soon. Thank you for

687
01:13:40,280 --> 01:13:47,000
listening to Artificially Intelligent Marketing. To stay on top of the latest trends, tips and tools

688
01:13:47,000 --> 01:14:04,280
in the world of marketing AI. Be sure to subscribe. We look forward to seeing you again next week.