1
00:00:00,000 --> 00:00:09,960
Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the

2
00:00:09,960 --> 00:00:15,700
latest trends, tips and tools in the world of marketing AI, helping you get the best

3
00:00:15,700 --> 00:00:18,540
results from your marketing efforts.

4
00:00:18,540 --> 00:00:23,240
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:23,240 --> 00:00:29,840
Hello everyone, welcome to episode 18 after our short summer hiatus.

6
00:00:29,840 --> 00:00:34,080
Apologies that we left you without all of your latest marketing and AI information that

7
00:00:34,080 --> 00:00:39,200
you need, but we are going to make up for it this week because we have a banger of two

8
00:00:39,200 --> 00:00:43,800
to three weeks of everything you need to know that happened in the world of marketing and

9
00:00:43,800 --> 00:00:44,800
AI.

10
00:00:44,800 --> 00:00:48,160
I'm here as usual with my very good friend, Martin Broadhurst.

11
00:00:48,160 --> 00:00:49,680
How are you, Martin?

12
00:00:49,680 --> 00:00:50,680
I'm good.

13
00:00:50,680 --> 00:00:55,760
I'm fully relaxed, although I'm a little bit throaty, a bit of a laryngitis.

14
00:00:55,760 --> 00:01:02,080
I don't know if you can hear that there, but maybe I partied too hard on my friend's wedding

15
00:01:02,080 --> 00:01:03,080
this weekend.

16
00:01:03,080 --> 00:01:07,200
Let's hope you recover quickly because you've got a busy couple of weeks ahead of you, haven't

17
00:01:07,200 --> 00:01:08,200
you?

18
00:01:08,200 --> 00:01:09,200
I have.

19
00:01:09,200 --> 00:01:10,200
Yeah.

20
00:01:10,200 --> 00:01:15,440
Tomorrow morning I fly out to Cleveland for the Marketing AI Institute Conference, MAKON,

21
00:01:15,440 --> 00:01:17,600
out in Cleveland, Ohio.

22
00:01:17,600 --> 00:01:21,720
Really looking forward to that, they've got some genuinely top-drawer speakers and I'm

23
00:01:21,720 --> 00:01:22,720
auto speaking.

24
00:01:22,720 --> 00:01:23,720
So that's nice.

25
00:01:23,720 --> 00:01:24,720
I didn't clear.

26
00:01:24,720 --> 00:01:30,640
I would have just said there's some top-drawer speakers, but I appreciate the humbleness

27
00:01:30,640 --> 00:01:31,640
there, Martin.

28
00:01:31,640 --> 00:01:34,720
So if you're going to be at that event, go and say hello to Martin because I'm sure he'd

29
00:01:34,720 --> 00:01:35,720
love to say hello to you.

30
00:01:35,720 --> 00:01:37,360
He'd love to hear your thoughts on the podcast.

31
00:01:37,360 --> 00:01:42,200
If you are lucky enough to be going to that event, it does look like an absolute stellar

32
00:01:42,200 --> 00:01:43,200
lineup.

33
00:01:43,200 --> 00:01:48,240
So Martin, I'm sure you'll come back and give us your top takeaways from the event when

34
00:01:48,240 --> 00:01:49,240
you return.

35
00:01:49,240 --> 00:01:50,240
Yeah, very much so.

36
00:01:50,240 --> 00:01:54,240
I'm hoping to get a few interviews with delegates and speakers while I'm out there.

37
00:01:54,240 --> 00:01:56,880
So a bit of content for the pod.

38
00:01:56,880 --> 00:01:57,880
Me likey.

39
00:01:57,880 --> 00:01:58,880
Right.

40
00:01:58,880 --> 00:01:59,880
There's a lot to get through.

41
00:01:59,880 --> 00:02:01,240
So we're going to jump through.

42
00:02:01,240 --> 00:02:05,960
Now usually we have a number of big stories that we look into and a bunch of short snippets,

43
00:02:05,960 --> 00:02:10,640
but there has been so much that's happened, dear listener, that we're just going to crack

44
00:02:10,640 --> 00:02:12,560
through at speed.

45
00:02:12,560 --> 00:02:15,960
We're going to try and take you through all the different things that have happened.

46
00:02:15,960 --> 00:02:21,560
And I'm sure Martin and I will jump into some conversation around some of the aspects where

47
00:02:21,560 --> 00:02:25,560
appropriate and there won't even be time for tool of the week this week.

48
00:02:25,560 --> 00:02:29,840
I suspect there'll be some tools of the week hidden in all of the updates because there's

49
00:02:29,840 --> 00:02:31,840
so much good stuff.

50
00:02:31,840 --> 00:02:32,840
Okay.

51
00:02:32,840 --> 00:02:34,740
Deep breath.

52
00:02:34,740 --> 00:02:36,720
Everybody get ready to drink from the fire hose.

53
00:02:36,720 --> 00:02:39,000
I'll get ready to turn it on.

54
00:02:39,000 --> 00:02:45,000
Let's start with a load of announcements this week from OpenAI, especially announcements

55
00:02:45,000 --> 00:02:46,000
around ChatGPT.

56
00:02:46,000 --> 00:02:50,840
So the first thing is a couple of weeks ago, according to the team at SimilarWeb, which

57
00:02:50,840 --> 00:02:55,560
is a platform that does its best to try and assess how much traffic different websites

58
00:02:55,560 --> 00:02:59,600
are getting, even though of course they don't have access to their actual data.

59
00:02:59,600 --> 00:03:04,840
But according to SimilarWeb, it looked like ChatGPT's user growth may have stalled because

60
00:03:04,840 --> 00:03:10,920
the data suggested that the platform had fewer users in June than in May.

61
00:03:10,920 --> 00:03:13,680
So that's quite an interesting one to kick us off with.

62
00:03:13,680 --> 00:03:19,320
Has the hype cycle of everybody wanting to have a play with ChatGPT, is it burning itself

63
00:03:19,320 --> 00:03:20,320
out?

64
00:03:20,320 --> 00:03:24,520
Yeah, I think we've probably saturated the early user growth.

65
00:03:24,520 --> 00:03:26,560
Now we're kind of heading into that.

66
00:03:26,560 --> 00:03:30,880
Okay, we've got the routine users, the people that have played with the free version and

67
00:03:30,880 --> 00:03:32,320
weren't blown away with it.

68
00:03:32,320 --> 00:03:35,000
We've now gone by the wayside.

69
00:03:35,000 --> 00:03:38,160
But you can actually see this reflected in Google Trends as well.

70
00:03:38,160 --> 00:03:42,920
If you look at Google Trends for ChatGPT, it has plateaued.

71
00:03:42,920 --> 00:03:46,640
So it's not just SimilarWeb that we can look to for this.

72
00:03:46,640 --> 00:03:49,560
It seems to be confirmed by multiple points.

73
00:03:49,560 --> 00:03:54,880
Yeah, and I've been trying to reflect on this during our little summer hiatus.

74
00:03:54,880 --> 00:03:59,200
And I think there's definitely anecdotally from people I've spoken with, people who've

75
00:03:59,200 --> 00:04:04,280
had a bit of a play, thought, oh, this is cool, but struggled to really apply it in

76
00:04:04,280 --> 00:04:05,280
their work.

77
00:04:05,280 --> 00:04:12,080
So I think adoption of AI is going to have to overcome a fairly major hump there when

78
00:04:12,080 --> 00:04:18,920
this is rolled out to the wider masses through Google Workspace and Microsoft 365 Copilot

79
00:04:18,920 --> 00:04:20,540
and these other tools.

80
00:04:20,540 --> 00:04:26,040
The other thing I was really thinking about is I had a conversation online with a couple

81
00:04:26,040 --> 00:04:32,560
of folks about Code Interpreter, which we'll talk about a bit later, and some of the analyses

82
00:04:32,560 --> 00:04:35,400
of Code Interpreter being wrong.

83
00:04:35,400 --> 00:04:40,680
And sometimes when you get outputs from ChatGPT, some of the stuff is made up.

84
00:04:40,680 --> 00:04:46,760
And it really made me think, how commercially viable are these tools if you can't trust

85
00:04:46,760 --> 00:04:47,760
them?

86
00:04:47,760 --> 00:04:50,040
Because it means you've got to check everything.

87
00:04:50,040 --> 00:04:55,800
Now I'm sure they're still able to save us time, even when you've got to do lots of

88
00:04:55,800 --> 00:04:59,400
checking, because at least you don't have to perhaps create the raw output yourself.

89
00:04:59,400 --> 00:05:06,960
But that does limit, I think, the impact that they can have and also how much quite a big

90
00:05:06,960 --> 00:05:11,160
chunk of the working populace is going to be willing to trust them with, right?

91
00:05:11,160 --> 00:05:15,240
Because they'll be worried that they're going to report back to their boss or other

92
00:05:15,240 --> 00:05:20,540
key figures in the business and outside of the business stuff that's wrong.

93
00:05:20,540 --> 00:05:24,040
So yeah, I think it's going to be interesting to actually see how that plays out.

94
00:05:24,040 --> 00:05:30,160
Yeah, I mean, imagine that you're a lawyer, for instance, going into a court case and

95
00:05:30,160 --> 00:05:36,400
citing examples from case law and none of them existing.

96
00:05:36,400 --> 00:05:37,400
Could you imagine?

97
00:05:37,400 --> 00:05:40,640
What a ridiculous human being you'd have to be to do that.

98
00:05:40,640 --> 00:05:41,840
But no, you're right.

99
00:05:41,840 --> 00:05:45,160
I think people are going to have to have trust in the systems.

100
00:05:45,160 --> 00:05:47,840
And actually that comes through exposure.

101
00:05:47,840 --> 00:05:52,560
The more you use the tools, the more you recognize where their strengths and weaknesses lie and

102
00:05:52,560 --> 00:05:57,120
how much you have to check and double check certain elements that it gives you.

103
00:05:57,120 --> 00:06:02,640
So I understand why people might be reticent to use them, particularly if they've only

104
00:06:02,640 --> 00:06:08,960
used the free version of ChatGPT with GPT 3.5.

105
00:06:08,960 --> 00:06:15,760
But once you start to see where it excels and how it can save you certain amounts of

106
00:06:15,760 --> 00:06:22,280
time on very specific tasks, like you probably have a set list of tasks that you know.

107
00:06:22,280 --> 00:06:28,000
ChatGPT will do quicker than you doing it on your own.

108
00:06:28,000 --> 00:06:32,960
As soon as you've done that, then you start to introduce it in your day to day.

109
00:06:32,960 --> 00:06:33,960
I think you're right.

110
00:06:33,960 --> 00:06:38,720
And I think as the tools get better, you can expand that list or start to introduce tasks

111
00:06:38,720 --> 00:06:41,520
the way previously you decided it wasn't quite good enough yet.

112
00:06:41,520 --> 00:06:43,840
Yeah, I think we're going to see that as well.

113
00:06:43,840 --> 00:06:51,040
And for those of you who are new to the show, the reference to the lawyer who used AI to

114
00:06:51,040 --> 00:06:54,800
drive a court case he was involved in is episode 13.

115
00:06:54,800 --> 00:06:56,480
Go back, check it out, see what we're talking about.

116
00:06:56,480 --> 00:06:58,280
This is kind of crazy.

117
00:06:58,280 --> 00:07:05,720
Moving swiftly on with the stories here then, OpenAI has also provided all API users, all

118
00:07:05,720 --> 00:07:11,720
is in quotation marks at the moment, with access to GPT 4.

119
00:07:11,720 --> 00:07:15,860
So this is a pretty big deal for those of you that are out there, moving beyond just

120
00:07:15,860 --> 00:07:21,480
playing with ChatGPT and trying to see if there are ways you can smart automate aspects

121
00:07:21,480 --> 00:07:23,760
of your work using tools like Zapier.

122
00:07:23,760 --> 00:07:30,920
And until now, most of us, including me still, which is annoying, had to rely on GPT model

123
00:07:30,920 --> 00:07:34,000
3.5, which is pretty good for lots of different stuff.

124
00:07:34,000 --> 00:07:37,680
But it doesn't have the content creation capabilities of GPT 4.

125
00:07:37,680 --> 00:07:43,360
So for example, Martin and I, and Martin spoke about an example that he did in a live event,

126
00:07:43,360 --> 00:07:46,040
which I can't even remember what episode that was, but you should definitely go back

127
00:07:46,040 --> 00:07:47,680
and listen to that one.

128
00:07:47,680 --> 00:07:52,920
And I've built myself an app where I can record voice notes on my phone and then the

129
00:07:52,920 --> 00:07:56,840
audio gets pushed into Whisper, which is a transcription tool.

130
00:07:56,840 --> 00:08:02,160
And then the transcription gets pushed into ChatGPT to be summarized and drafted into

131
00:08:02,160 --> 00:08:03,640
an email that I can send to someone.

132
00:08:03,640 --> 00:08:05,560
But they all rely on GPT 3.5.

133
00:08:05,560 --> 00:08:07,400
And so the quality is a bit meh.

134
00:08:07,400 --> 00:08:11,600
But if you can get access to GPT 4, it's going to seriously upgrade what you can do,

135
00:08:11,600 --> 00:08:14,160
building those self automations.

136
00:08:14,160 --> 00:08:18,400
But so I think that's really cool if you're in that niche power group.

137
00:08:18,400 --> 00:08:21,720
Obviously we're going to see even more tools improve a lot if everybody can access GPT

138
00:08:21,720 --> 00:08:23,200
4 via the API.

139
00:08:23,200 --> 00:08:25,040
Have you got access yet, Martin?

140
00:08:25,040 --> 00:08:26,040
I have yet.

141
00:08:26,040 --> 00:08:31,480
And I've made some upgrades to some of the workflows that I had in Zapier.

142
00:08:31,480 --> 00:08:33,520
I was up here.

143
00:08:33,520 --> 00:08:36,600
Yet that's, I've seen some improvements.

144
00:08:36,600 --> 00:08:40,560
The prompts that I had anyway gave me pretty good outputs with 3.5.

145
00:08:40,560 --> 00:08:45,000
So I haven't seen any dramatic changes except the cost.

146
00:08:45,000 --> 00:08:50,240
I mean, I'm using it at a very low level, but the price difference between 3.5 and 4

147
00:08:50,240 --> 00:08:52,440
is, yeah, it's noticeable.

148
00:08:52,440 --> 00:08:53,440
Interesting.

149
00:08:53,440 --> 00:08:57,400
So you may not be getting that much better results, but you are paying more.

150
00:08:57,400 --> 00:09:02,240
It's going to be, it's definitely going to be used per case by use case, isn't it?

151
00:09:02,240 --> 00:09:03,240
Without a doubt.

152
00:09:03,240 --> 00:09:06,800
And that's the thing I think people need to recognize is as a clammer to get access to

153
00:09:06,800 --> 00:09:08,840
something like GPT 4.

154
00:09:08,840 --> 00:09:13,720
And actually for lots of use cases, particularly with good prompting and few shot and many

155
00:09:13,720 --> 00:09:20,400
shot prompting as well, you can get really good results with 3.5.

156
00:09:20,400 --> 00:09:21,400
There we are.

157
00:09:21,400 --> 00:09:23,160
Marketers playing with Zapier and other tools.

158
00:09:23,160 --> 00:09:27,320
Bit of advice there on how to perhaps get the best bang for buck by choosing the right

159
00:09:27,320 --> 00:09:30,000
model for the right application.

160
00:09:30,000 --> 00:09:32,180
Some more ChatGPT news.

161
00:09:32,180 --> 00:09:39,500
Many of you would have noticed this actually by now, but ChatGPT has seen its internet

162
00:09:39,500 --> 00:09:45,000
access revoked as users were reportedly using it to get around paywalls and access gated

163
00:09:45,000 --> 00:09:47,960
content on top news sites.

164
00:09:47,960 --> 00:09:53,240
And as of today, internet access for ChatGPT hasn't yet to be reinstated.

165
00:09:53,240 --> 00:10:00,480
Yeah, that was a, it was a use case that people found really early on as well.

166
00:10:00,480 --> 00:10:03,920
So I'm surprised it took them so long to pull it.

167
00:10:03,920 --> 00:10:11,440
But yeah, I'm not surprised that publishers were up in arms about that.

168
00:10:11,440 --> 00:10:15,040
No doubt we'll see it reinstated fairly soon.

169
00:10:15,040 --> 00:10:20,600
In the meantime, you can always use the plugin WebPilot if you want to browse the web.

170
00:10:20,600 --> 00:10:26,480
It doesn't have quite the same capabilities as the Bing one, but for most users, they

171
00:10:26,480 --> 00:10:28,360
probably won't notice the differences.

172
00:10:28,360 --> 00:10:34,000
So yeah, if you do want to use ChatGPT connected to the web, just use the plugin WebPilot.

173
00:10:34,000 --> 00:10:36,620
That's a good bit of advice.

174
00:10:36,620 --> 00:10:39,440
You can also use Bing.

175
00:10:39,440 --> 00:10:43,480
And if you're feeling particularly brave, you can use Bard.

176
00:10:43,480 --> 00:10:47,560
But we'll talk about the, there's been some improvements to Bard, but Bard is still very

177
00:10:47,560 --> 00:10:53,600
much lagging behind ChatGPT in a lot of, a lot of applications these days.

178
00:10:53,600 --> 00:10:55,700
Right.

179
00:10:55,700 --> 00:10:59,960
First is from OpenAI that they've announced that they've tasked a team internally with

180
00:10:59,960 --> 00:11:04,680
the goal of building a human level automated alignment researcher, which can be used to

181
00:11:04,680 --> 00:11:08,280
scale efforts and help align super intelligence.

182
00:11:08,280 --> 00:11:14,160
So on the one hand, this appears to be a proactive step to sort of help quell those fears around

183
00:11:14,160 --> 00:11:18,520
powerful AI tools coming out over the next few months and years that become less and

184
00:11:18,520 --> 00:11:20,440
less aligned with human values.

185
00:11:20,440 --> 00:11:24,360
But one assumes it will also help OpenAI to align their models without having to rely

186
00:11:24,360 --> 00:11:29,960
so much on learning with reinforcement, learning with human feedback, which obviously takes

187
00:11:29,960 --> 00:11:31,420
a lot of people and a lot of time.

188
00:11:31,420 --> 00:11:37,360
So there's a commercial benefit to creating this, a fairly massive one, one would have

189
00:11:37,360 --> 00:11:38,360
thought.

190
00:11:38,360 --> 00:11:42,120
But most of the story from OpenAI has been about spinning.

191
00:11:42,120 --> 00:11:47,520
They're on a voyage to make sure that AGI remains aligned with what humans want, not

192
00:11:47,520 --> 00:11:48,720
just what it wants.

193
00:11:48,720 --> 00:11:51,760
So I thought it was quite an interesting story from that perspective.

194
00:11:51,760 --> 00:11:56,480
It's a great marketing tool, isn't it, for the company as well.

195
00:11:56,480 --> 00:12:05,720
When you say, look, we're investing so much into keeping our incredibly powerful artificial

196
00:12:05,720 --> 00:12:11,240
intelligence system aligned to humans because our system, which is incredibly powerful and

197
00:12:11,240 --> 00:12:15,980
incredibly capable and can do all of these amazing things, we have to invest heavily

198
00:12:15,980 --> 00:12:16,980
to keep it safe.

199
00:12:16,980 --> 00:12:20,040
By the way, did we mention that our system is incredibly powerful?

200
00:12:20,040 --> 00:12:26,800
It's a great marketing tool when you build this expectation and make people believe that

201
00:12:26,800 --> 00:12:33,400
you've got this product which is capable of so much, potentially so dangerous, potentially

202
00:12:33,400 --> 00:12:35,000
being the key word.

203
00:12:35,000 --> 00:12:37,920
It's got some wonderful marketing 101 in it, hasn't it?

204
00:12:37,920 --> 00:12:41,320
Choose the one, two or three key features and benefits that you're going to promote

205
00:12:41,320 --> 00:12:45,640
and repeat yourself a lot until it's embedded in people's brains.

206
00:12:45,640 --> 00:12:49,840
So yeah, that was a cool one.

207
00:12:49,840 --> 00:12:55,200
Not as cool as our next bit of news, which is that OpenAI made code interpreter available

208
00:12:55,200 --> 00:12:56,560
for everyone.

209
00:12:56,560 --> 00:13:02,600
So this was in a closed alpha, I think, and now you can access it via beta, but it's kind

210
00:13:02,600 --> 00:13:03,600
of awesome.

211
00:13:03,600 --> 00:13:04,600
In fact, it's so powerful.

212
00:13:04,600 --> 00:13:09,240
There's been a theory going around that in essence, you could consider code interpreter

213
00:13:09,240 --> 00:13:16,960
as GPT 4.5 because I think what a lot of people know it for is its abilities with data analysis.

214
00:13:16,960 --> 00:13:19,040
So we should probably touch on those first, Martin.

215
00:13:19,040 --> 00:13:24,480
So within just a chat with code interpreter plugin enabled in ChatGPT, you can do data

216
00:13:24,480 --> 00:13:27,800
analysis, data visualization, predictive modeling.

217
00:13:27,800 --> 00:13:29,280
You can clean your data up.

218
00:13:29,280 --> 00:13:31,800
You can even create synthetic data.

219
00:13:31,800 --> 00:13:36,440
So they're all quite well-known use cases and I've had a bit of a play with those.

220
00:13:36,440 --> 00:13:38,360
We can look at those in a minute.

221
00:13:38,360 --> 00:13:42,800
But because it can produce its own Python code, you can also get it to do other interesting

222
00:13:42,800 --> 00:13:49,440
things for you like manipulating audio files, creating PDFs, creating CSV files.

223
00:13:49,440 --> 00:13:54,160
So basically, you could give it some messy data and then ask it to clean it up and send

224
00:13:54,160 --> 00:13:58,800
it back to you and you download the CSV file.

225
00:13:58,800 --> 00:14:04,160
Some people have even used it to create interactive dashboards and interactive maps, which is really

226
00:14:04,160 --> 00:14:05,160
awesome.

227
00:14:05,160 --> 00:14:10,400
Obviously not in the chat window itself because it can't run the code that it produces, but

228
00:14:10,400 --> 00:14:16,200
you don't need to be a Python coder to be able to use the code.

229
00:14:16,200 --> 00:14:20,320
You just need to ask what you want and then it gives it back to you.

230
00:14:20,320 --> 00:14:22,280
So it's pretty amazing.

231
00:14:22,280 --> 00:14:25,720
There was one example where it looked like it could uncover and detect the number of

232
00:14:25,720 --> 00:14:27,560
faces in an image.

233
00:14:27,560 --> 00:14:31,200
So it's like, what is the limits of code interpreter?

234
00:14:31,200 --> 00:14:34,880
Because if you thought it was just going to be something to chuck some data in and make

235
00:14:34,880 --> 00:14:39,720
some nice graphs or ask some questions of your data, it can clearly do a lot more than

236
00:14:39,720 --> 00:14:40,720
that.

237
00:14:40,720 --> 00:14:52,640
Is it the early preview to the multimodality capabilities of GPT-4 that we saw previously?

238
00:14:52,640 --> 00:14:58,400
You mentioned something there about the data cleansing and being able to throw files in

239
00:14:58,400 --> 00:15:00,400
and it will export some.

240
00:15:00,400 --> 00:15:02,600
Have you tried it?

241
00:15:02,600 --> 00:15:04,000
Not for cleansing, have you?

242
00:15:04,000 --> 00:15:05,000
Yeah.

243
00:15:05,000 --> 00:15:11,760
It was the first use case that I put together and I created some kind of dummy data using

244
00:15:11,760 --> 00:15:15,880
an export of HubSpot contact sheet.

245
00:15:15,880 --> 00:15:22,440
So I just randomly capitalized and lower cased some names and put some missing data in fields

246
00:15:22,440 --> 00:15:26,120
like country and city and things like that.

247
00:15:26,120 --> 00:15:29,480
Made the inconsistent telephone numbers, that kind of thing.

248
00:15:29,480 --> 00:15:32,080
I just made it a little bit untidy.

249
00:15:32,080 --> 00:15:36,880
And it was only about 30 records in this sample document.

250
00:15:36,880 --> 00:15:41,320
And could I get code interpreter to fix it and give me the output?

251
00:15:41,320 --> 00:15:42,320
No, I could not.

252
00:15:42,320 --> 00:15:43,320
It was a CSV file.

253
00:15:43,320 --> 00:15:45,480
I stuck it in there, asked it to do it.

254
00:15:45,480 --> 00:15:55,600
It went through lots of creating Python, but it never got me the output that I wanted.

255
00:15:55,600 --> 00:15:58,320
Maybe this is bad prompting from my part.

256
00:15:58,320 --> 00:16:04,920
So then I thought, well, okay, how easy is it to just get it to do it in GPT-4, just

257
00:16:04,920 --> 00:16:07,600
copying and pasting the CSV?

258
00:16:07,600 --> 00:16:08,600
The data in, yeah.

259
00:16:08,600 --> 00:16:09,880
Yeah, the data in.

260
00:16:09,880 --> 00:16:17,320
Did that with my instructions in the prompt, got the output first time.

261
00:16:17,320 --> 00:16:20,280
But this harks back to the very first point in today's podcast, right?

262
00:16:20,280 --> 00:16:24,960
Which is like figuring out the horses for courses, like what's the right use case for

263
00:16:24,960 --> 00:16:29,720
each tool, but crumbs is a lot of effort to go through and try and figure out how to get

264
00:16:29,720 --> 00:16:32,920
it to make it work and then when to give up.

265
00:16:32,920 --> 00:16:34,920
Yeah, I'm on board.

266
00:16:34,920 --> 00:16:41,120
My main use case so far was I tried to think about what data sources I had with like a

267
00:16:41,120 --> 00:16:42,120
ton of data.

268
00:16:42,120 --> 00:16:45,080
Like I really wanted to like try and break it.

269
00:16:45,080 --> 00:16:51,080
So I exported a month's worth of time entry data from across the agency and I removed

270
00:16:51,080 --> 00:16:58,000
any client specific stuff so it was literally just people and time and date of timestamp.

271
00:16:58,000 --> 00:17:01,200
And I got it to make some really nice graphs for me.

272
00:17:01,200 --> 00:17:05,980
I wanted it to describe the data because I'd seen some examples online where you give it

273
00:17:05,980 --> 00:17:09,320
some data and then you say, what do you think are the 10 most interesting trends from this

274
00:17:09,320 --> 00:17:11,080
data?

275
00:17:11,080 --> 00:17:14,360
I got it to graph things like who'd done the most billable time.

276
00:17:14,360 --> 00:17:18,240
I then asked it in a text prompt, who's done the most billable time?

277
00:17:18,240 --> 00:17:23,760
And even though it previously just produced a graph for me that showed that person A had,

278
00:17:23,760 --> 00:17:27,520
it swore blind that someone completely different had done the most time.

279
00:17:27,520 --> 00:17:28,640
And I was like, are you sure?

280
00:17:28,640 --> 00:17:30,080
Because I'm pretty sure it's this person.

281
00:17:30,080 --> 00:17:33,420
And then the code interpreter went, oh no, yeah, you are right.

282
00:17:33,420 --> 00:17:35,280
And I was like, who's the second most?

283
00:17:35,280 --> 00:17:40,360
And it gave me the person that had given me before who wasn't even in the top five.

284
00:17:40,360 --> 00:17:44,520
And so I got into not a debate, but like an interesting conversation online about this

285
00:17:44,520 --> 00:17:50,360
where other people weren't having these problems, but when I've like tried three things and

286
00:17:50,360 --> 00:17:54,880
two of them are wrong, that's not the type of success rate where I feel like I can trust

287
00:17:54,880 --> 00:17:56,520
anything that comes out.

288
00:17:56,520 --> 00:17:58,140
Not even close.

289
00:17:58,140 --> 00:18:00,860
So who's this for?

290
00:18:00,860 --> 00:18:02,240
Because you know what?

291
00:18:02,240 --> 00:18:03,840
I'll chuck it in a pivot table, thanks.

292
00:18:03,840 --> 00:18:06,200
Cause I know the pivot table is going to be right.

293
00:18:06,200 --> 00:18:10,920
And I appreciate not everybody knows how to create pivot tables or often people are not

294
00:18:10,920 --> 00:18:16,560
working in them enough to remember their functionality or to really feel comfortable using them.

295
00:18:16,560 --> 00:18:20,680
But at least the information's correct.

296
00:18:20,680 --> 00:18:25,100
If I have to do a pivot table and the code interpreter analysis and double check them

297
00:18:25,100 --> 00:18:29,400
against each other, I'll skip the code interpreter, right?

298
00:18:29,400 --> 00:18:31,800
Yeah, absolutely.

299
00:18:31,800 --> 00:18:36,040
The more I think about this, the more I'm like, there's going to be a bunch of use cases

300
00:18:36,040 --> 00:18:40,440
that I'm just not going to touch until there's some mechanisms in place to ensure quality

301
00:18:40,440 --> 00:18:42,760
of output and accuracy.

302
00:18:42,760 --> 00:18:43,760
Right?

303
00:18:43,760 --> 00:18:50,920
And I think that we've talked about it a lot on the part about UI, UX is so important and

304
00:18:50,920 --> 00:18:52,080
you've got to get that right.

305
00:18:52,080 --> 00:18:56,360
And what you want is a system where you've removed the need for prompting, right?

306
00:18:56,360 --> 00:18:59,760
Because prompting is half of the issue here.

307
00:18:59,760 --> 00:19:02,520
You put in one prompt, it doesn't quite work.

308
00:19:02,520 --> 00:19:03,520
So you're tweaking it.

309
00:19:03,520 --> 00:19:04,520
You constantly go back and forth.

310
00:19:04,520 --> 00:19:08,000
You just want to be able to stick your data in and get something interesting out.

311
00:19:08,000 --> 00:19:15,080
And this is where Copilot and the whatever the Google AI version for workspace is called.

312
00:19:15,080 --> 00:19:17,160
That's the nut that they're going to have to crack, isn't it?

313
00:19:17,160 --> 00:19:21,320
You hope that they can do this at the front end and it's not going to leave users having

314
00:19:21,320 --> 00:19:26,020
to go back and forth constantly prompting and reprompting in order to get something

315
00:19:26,020 --> 00:19:27,480
functional out of it.

316
00:19:27,480 --> 00:19:29,440
Yeah, I think that's definitely an issue.

317
00:19:29,440 --> 00:19:35,240
But I'm also thinking if I asked ChatGVT to create some content and it makes up scientific

318
00:19:35,240 --> 00:19:41,640
citations or whatever that sound plausible because of how its algorithms produce content

319
00:19:41,640 --> 00:19:45,840
based on its training data, I can sort of understand that.

320
00:19:45,840 --> 00:19:51,340
But when I give it some data, it shows me a graph output that clearly summarizes the

321
00:19:51,340 --> 00:19:54,880
data with a clear sort of outcome to the answer to the question.

322
00:19:54,880 --> 00:19:59,600
And the question was not ambiguous and it still gets it wrong.

323
00:19:59,600 --> 00:20:03,180
That's a worry for me because that's the biggest indicator yet.

324
00:20:03,180 --> 00:20:06,800
And I know that this is true, but it's hard when you're working with ChatGVT and you see

325
00:20:06,800 --> 00:20:09,040
it output and you think, God, it knows stuff.

326
00:20:09,040 --> 00:20:12,840
But that was the biggest indicator to me yet that it knows nothing.

327
00:20:12,840 --> 00:20:19,600
It's just somehow regurgitating information based on other information.

328
00:20:19,600 --> 00:20:25,320
I'm surprised Code Interpreter works as well as it does, to be honest, but until I can

329
00:20:25,320 --> 00:20:28,880
really trust it, I'm not sure I can make it part of my workflow.

330
00:20:28,880 --> 00:20:33,080
Not to mention the fact I had to anonymize a load of the data because they couldn't risk

331
00:20:33,080 --> 00:20:34,960
putting any client specific info in.

332
00:20:34,960 --> 00:20:39,240
So I can't ask anything like, oh, what's the profitability of this client versus this client

333
00:20:39,240 --> 00:20:40,240
or this project?

334
00:20:40,240 --> 00:20:43,720
Because I just can't trust it to give it the data.

335
00:20:43,720 --> 00:20:46,480
So limitations abound.

336
00:20:46,480 --> 00:20:51,160
People are doing some cool stuff with it, as we'll look at a little bit later.

337
00:20:51,160 --> 00:20:55,200
But I think it's going to take a fair bit personally to realize its potential, certainly

338
00:20:55,200 --> 00:20:59,200
in our hands, Martin, because we can't get it to do even basic things that we want, which

339
00:20:59,200 --> 00:21:03,880
either means that we are not the brightest, which is highly plausible.

340
00:21:03,880 --> 00:21:07,160
Maybe we'll ask Code Interpreter its opinion on that.

341
00:21:07,160 --> 00:21:11,200
Or it's not easy to get correct answers out of it.

342
00:21:11,200 --> 00:21:13,160
Right, moving on.

343
00:21:13,160 --> 00:21:16,480
Still on the OpenAI, still on chunk one here, all your news.

344
00:21:16,480 --> 00:21:20,500
OpenAI is looking for new sources of data for its model, striking deals with the likes

345
00:21:20,500 --> 00:21:26,140
of Shutterstock to access its images, videos, music and metadata, and also the Associated

346
00:21:26,140 --> 00:21:30,080
Press tapping into its news archives dating back to 1985.

347
00:21:30,080 --> 00:21:32,320
So there's an interesting one.

348
00:21:32,320 --> 00:21:37,520
The ongoing pursuit of information to train your models on and how to get it.

349
00:21:37,520 --> 00:21:42,080
I couldn't help the cynic in me, Martin, couldn't help but think, have they accessed this already

350
00:21:42,080 --> 00:21:45,360
and they're just sweeping up behind themselves?

351
00:21:45,360 --> 00:21:48,600
Yeah, that's a...

352
00:21:48,600 --> 00:21:50,960
What's the phrase?

353
00:21:50,960 --> 00:21:55,680
It's easier to ask for forgiveness or beg forgiveness than ask permission.

354
00:21:55,680 --> 00:22:00,600
Well, yeah, and you've got to wonder, especially you've got the likes of Sarah Silverman suing

355
00:22:00,600 --> 00:22:05,640
at the moment because ChatGPT is just too good at summarizing books to have not been

356
00:22:05,640 --> 00:22:08,440
showing the original information and stuff like that.

357
00:22:08,440 --> 00:22:10,720
I think we could see a lot of this.

358
00:22:10,720 --> 00:22:16,960
At some point, we're going to have our Napster to Spotify transition, right, from the Wild

359
00:22:16,960 --> 00:22:23,960
Wild West, everything's fair game, everything's free to somehow we have to at least partially

360
00:22:23,960 --> 00:22:29,760
compensate people who helped create the system, i.e. all the content creators.

361
00:22:29,760 --> 00:22:35,240
I'm glad that's not my job to figure out how to fix that, but I suspect we're going to

362
00:22:35,240 --> 00:22:36,240
have to get there.

363
00:22:36,240 --> 00:22:40,160
Right, moving on, ChatGPT is getting worse, question mark.

364
00:22:40,160 --> 00:22:41,160
Perhaps, perhaps not.

365
00:22:41,160 --> 00:22:46,000
So there was a new paper out by researchers at Stanford and Berkeley that suggested that

366
00:22:46,000 --> 00:22:51,920
GPT 3.5 and GPT 4 have changed a lot in their performance over a short period.

367
00:22:51,920 --> 00:22:56,200
The researchers evaluate ChatGPT on various tasks such as solving math problems, answering

368
00:22:56,200 --> 00:23:00,080
sensitive questions, generating code and reasoning visually.

369
00:23:00,080 --> 00:23:04,800
And what their data showed is that ChatGPT had become much worse at some tasks, so things

370
00:23:04,800 --> 00:23:12,960
like identifying prime numbers, also coding, but that GPT 3.5 bizarrely had got better

371
00:23:12,960 --> 00:23:14,800
at some tasks.

372
00:23:14,800 --> 00:23:21,020
Now other researchers have massively questioned the approach that this paper took and so the

373
00:23:21,020 --> 00:23:25,200
results might have to be taken with quite a large dose of salt.

374
00:23:25,200 --> 00:23:29,120
There's also some suggestions that the more you use GPT 4 and you start to realise where

375
00:23:29,120 --> 00:23:33,480
it doesn't work very well, the more that becomes obvious to you and it's your perception of

376
00:23:33,480 --> 00:23:37,320
GPT 4 that's changing, not GPT 4 itself.

377
00:23:37,320 --> 00:23:41,680
That being said, if you're a user of ChatGPT and you have felt recently like its performance

378
00:23:41,680 --> 00:23:46,920
might have been waning, there may be at least some data to suggest that's true.

379
00:23:46,920 --> 00:23:49,400
Thoughts on this Martin?

380
00:23:49,400 --> 00:23:53,560
I suspect it's more a case of familiarity.

381
00:23:53,560 --> 00:23:59,640
I saw the feedback from other researchers questioning the approach.

382
00:23:59,640 --> 00:24:05,040
The Twitter threads that I read seem to make some good points, although unfortunately for

383
00:24:05,040 --> 00:24:08,000
listeners I can't remember those points.

384
00:24:08,000 --> 00:24:09,720
But yeah, I think it's familiarity.

385
00:24:09,720 --> 00:24:15,160
I think you just, much like knowing the limitations we talked about earlier on, knowing the limitations

386
00:24:15,160 --> 00:24:19,240
of a system so therefore you don't try to get it to do something or you don't publish

387
00:24:19,240 --> 00:24:24,560
content that's hallucinated or something, you also recognise the strengths more quickly.

388
00:24:24,560 --> 00:24:30,440
And when you've gone from 3.5 to 4, when they first announced 4 and we all started playing

389
00:24:30,440 --> 00:24:35,120
with GPT 4, you immediately went, oh my God, this is a step change.

390
00:24:35,120 --> 00:24:40,800
But then once you've ran a couple of hundred, couple of thousand prompts through it, you

391
00:24:40,800 --> 00:24:45,040
start going, hmm, it's not all that, is it?

392
00:24:45,040 --> 00:24:46,040
Just say some silly things.

393
00:24:46,040 --> 00:24:49,360
So I think it's more about familiarity.

394
00:24:49,360 --> 00:24:51,560
I think that's definitely part of it.

395
00:24:51,560 --> 00:24:57,480
It's interesting and I'd have to dive deeper into the paper to see how robust those assessments

396
00:24:57,480 --> 00:25:01,360
were that showed that it was good at one thing before and not so good now.

397
00:25:01,360 --> 00:25:07,960
I really question the prime number thing.

398
00:25:07,960 --> 00:25:15,080
The 97.6% accuracy reduced to 2.5%.

399
00:25:15,080 --> 00:25:18,360
I don't know.

400
00:25:18,360 --> 00:25:23,040
Maybe they have significantly changed the weights and biases in the model such that they've

401
00:25:23,040 --> 00:25:26,160
just absolutely kneecapped it at maths.

402
00:25:26,160 --> 00:25:30,760
But it would seem that that's too dramatic, isn't it?

403
00:25:30,760 --> 00:25:31,760
Surely?

404
00:25:31,760 --> 00:25:32,760
I mean, you would guess so.

405
00:25:32,760 --> 00:25:36,360
I think for me, it's more about how robust was the assessment method unless they knew

406
00:25:36,360 --> 00:25:40,360
for a fact that in three months time that we're going to run the same analysis, which

407
00:25:40,360 --> 00:25:42,860
would be a smart experiment.

408
00:25:42,860 --> 00:25:45,360
How did the conditions change?

409
00:25:45,360 --> 00:25:46,360
I don't know.

410
00:25:46,360 --> 00:25:50,480
But ultimately, we're talking about learning what these tools are good at and not good

411
00:25:50,480 --> 00:25:51,480
at.

412
00:25:51,480 --> 00:25:56,960
And if they can change quite a lot, that's going to make that harder because you're going

413
00:25:56,960 --> 00:26:01,840
to be like, oh, okay, I can rely on that for this.

414
00:26:01,840 --> 00:26:05,360
But yet, we're talking a lot about how the models need to improve.

415
00:26:05,360 --> 00:26:08,040
If you want the models to improve, then by definition, they're going to have to change.

416
00:26:08,040 --> 00:26:10,960
They're probably going to get better at some things and worse than others.

417
00:26:10,960 --> 00:26:14,200
How do you do that in an environment where people are trying to learn to use the tool,

418
00:26:14,200 --> 00:26:15,760
but it's not static?

419
00:26:15,760 --> 00:26:20,600
Gmail, to all intents and purposes, is a static tool.

420
00:26:20,600 --> 00:26:24,480
It doesn't change how it sends emails and I have to figure out a new button to press

421
00:26:24,480 --> 00:26:25,760
next week.

422
00:26:25,760 --> 00:26:27,720
And this is important for the adoption.

423
00:26:27,720 --> 00:26:30,880
You talk a lot about UX and I think you bang on mine.

424
00:26:30,880 --> 00:26:35,240
It's important for the adoption of tools that you know how to get the things out of them

425
00:26:35,240 --> 00:26:38,040
that you need and what you can rely on them for.

426
00:26:38,040 --> 00:26:44,440
And I think if it ends up being too dynamic, that could be a challenge.

427
00:26:44,440 --> 00:26:49,080
I appreciate pressing a button in Gmail and pressing a button in ChatGPT from a UX perspective

428
00:26:49,080 --> 00:26:50,360
are basically the same thing.

429
00:26:50,360 --> 00:26:55,000
But I mean, the full user experience, not just the visual user experience.

430
00:26:55,000 --> 00:26:56,720
So right, we better keep moving.

431
00:26:56,720 --> 00:26:58,360
There's still loads to go.

432
00:26:58,360 --> 00:26:59,360
Right.

433
00:26:59,360 --> 00:27:03,240
ChatGPT is increasing message limits for those of you that are interested in this.

434
00:27:03,240 --> 00:27:06,600
I personally encourage it at 50 messages per three hours.

435
00:27:06,600 --> 00:27:09,160
Those have reported at being 100.

436
00:27:09,160 --> 00:27:14,600
In all cases, it looks like this is an improvement over the 25 message cap per three hours, which

437
00:27:14,600 --> 00:27:18,540
I know a lot of people were hitting and it is quite annoying.

438
00:27:18,540 --> 00:27:20,280
So that's great.

439
00:27:20,280 --> 00:27:27,160
Some more news, especially if you're not in Europe or the UK, is that ChatGPT and OpenI

440
00:27:27,160 --> 00:27:28,960
have introduced custom instructions.

441
00:27:28,960 --> 00:27:31,120
So this is really quite cool.

442
00:27:31,120 --> 00:27:35,000
Basically, this is like introducing custom prompt injections for every chat that you

443
00:27:35,000 --> 00:27:40,920
have because you provide in your settings some information about you, how you would

444
00:27:40,920 --> 00:27:50,040
like ChatGPT to respond and any sort of delimitations or context for basically all responses you'll

445
00:27:50,040 --> 00:27:51,520
ever need.

446
00:27:51,520 --> 00:27:56,040
We've talked previously about getting it to imitate your email style.

447
00:27:56,040 --> 00:28:01,400
One would assume if you include a bit of information about your writing style as part of your custom

448
00:28:01,400 --> 00:28:05,160
instructions that would make that all work a little bit better.

449
00:28:05,160 --> 00:28:09,680
Now, as I mentioned, you can't access this in the UK and EU at the moment, although you

450
00:28:09,680 --> 00:28:13,120
can with a VPN, so you can have a play.

451
00:28:13,120 --> 00:28:18,560
So you can't test it if you're in where we are right now, not easily anyway, but initial

452
00:28:18,560 --> 00:28:22,760
users are praising its ability to generate better outputs without having to constantly

453
00:28:22,760 --> 00:28:26,800
enter a ton of the same prompt info, which is nice.

454
00:28:26,800 --> 00:28:30,160
There's also been some incredible creative users of this tool.

455
00:28:30,160 --> 00:28:35,560
So a person called Nick Dobos on Twitter was suggesting you can use it to build full agents,

456
00:28:35,560 --> 00:28:38,040
a bit like AutoGBT or BabyGBT.

457
00:28:38,040 --> 00:28:43,640
In essence, you use Code interpreter, you get it to create a list of tasks that it then

458
00:28:43,640 --> 00:28:49,720
saves to a file, and then you ask it to pull the tasks from that file one by one to take

459
00:28:49,720 --> 00:28:52,840
you through a process, which was quite interesting.

460
00:28:52,840 --> 00:28:57,880
Nick also found a way of, in effect, having it summarize a bit like a choose your own

461
00:28:57,880 --> 00:29:03,280
adventure, but for your own logical analysis, the options that you might want to add and

462
00:29:03,280 --> 00:29:04,960
with hotkeys.

463
00:29:04,960 --> 00:29:09,360
So if you want this, press one, if you want this, press two, if you want this, press three.

464
00:29:09,360 --> 00:29:13,720
So you could actually cycle through it quite quickly without actually even having to type,

465
00:29:13,720 --> 00:29:18,400
which then ended up being a really good example of how you can use this custom instructions

466
00:29:18,400 --> 00:29:22,560
to build that system and then Code interpreter to do some cool stuff that's got nothing to

467
00:29:22,560 --> 00:29:24,380
do in essence with data.

468
00:29:24,380 --> 00:29:26,380
So custom instructions.

469
00:29:26,380 --> 00:29:27,380
Looks like it could be powerful.

470
00:29:27,380 --> 00:29:29,600
Hopefully they'll roll out wider use soon.

471
00:29:29,600 --> 00:29:32,640
What are your thoughts on custom instructions, Martin?

472
00:29:32,640 --> 00:29:34,480
I'm looking forward to getting my hands on with it.

473
00:29:34,480 --> 00:29:39,040
I forgot that it wasn't available in the UK because I did have access to it briefly and

474
00:29:39,040 --> 00:29:42,080
I must have had my VPN on or something.

475
00:29:42,080 --> 00:29:45,400
And then I went onto it this morning and I've just posted on threads.

476
00:29:45,400 --> 00:29:46,400
Where has it gone?

477
00:29:46,400 --> 00:29:50,840
So yeah, I'm glad you've reminded me that that's why I can't see it.

478
00:29:50,840 --> 00:29:52,760
But it's a cool feature.

479
00:29:52,760 --> 00:29:59,440
And I've seen a few users giving some use cases that I'm absolutely going to use.

480
00:29:59,440 --> 00:30:06,180
For example, just Sam Altman himself posted one on Twitter about how he's using it, which

481
00:30:06,180 --> 00:30:10,280
is to basically say, I know you're a large language model.

482
00:30:10,280 --> 00:30:15,120
You do not need to tell me every time, just give me the answer.

483
00:30:15,120 --> 00:30:16,960
And yet little things like that.

484
00:30:16,960 --> 00:30:23,520
The amount of times I have to do this really basic jailbreak where I'm saying, I know you're

485
00:30:23,520 --> 00:30:30,600
an AI, I'm a human and yes, I will seek professional advice and blah, blah, blah.

486
00:30:30,600 --> 00:30:32,400
But just role play with me.

487
00:30:32,400 --> 00:30:35,200
If I can get rid of all of that, then great.

488
00:30:35,200 --> 00:30:37,320
I've saved some time.

489
00:30:37,320 --> 00:30:42,600
So my favorite thing about that is it's quite similar to what Nick does in the example I

490
00:30:42,600 --> 00:30:43,680
just mentioned.

491
00:30:43,680 --> 00:30:50,320
The top of his injection is no chat, just do.

492
00:30:50,320 --> 00:30:54,040
No chat semicolon, just do.

493
00:30:54,040 --> 00:31:01,880
That feels like if you want to be the open AI competitor in the style of Nike, Ikea I

494
00:31:01,880 --> 00:31:06,240
should say, that's probably how you have to frame your strap line.

495
00:31:06,240 --> 00:31:07,240
Right.

496
00:31:07,240 --> 00:31:13,160
Let's talk about Claw 2 because the ridiculous generative AI text generation news doesn't

497
00:31:13,160 --> 00:31:14,160
help.

498
00:31:14,160 --> 00:31:18,880
So Anthropic, who we've talked about on the podcast a fair amount, have released the next

499
00:31:18,880 --> 00:31:23,880
version of their Claude model and it is a significant leap forward.

500
00:31:23,880 --> 00:31:29,440
So it's nearly as good as ChatGPT 4 on a variety of standard benchmark tests.

501
00:31:29,440 --> 00:31:31,840
It has a 100K context window.

502
00:31:31,840 --> 00:31:32,840
We've talked about this before.

503
00:31:32,840 --> 00:31:33,840
So basically about 75,000 words.

504
00:31:33,840 --> 00:31:40,180
So it can handle a lot more input than ChatGPT, which makes it brilliant for summarizing large

505
00:31:40,180 --> 00:31:41,760
amounts of info.

506
00:31:41,760 --> 00:31:48,240
They've also changed how the UX works.

507
00:31:48,240 --> 00:31:51,480
So it's much more chat-like and easier to use.

508
00:31:51,480 --> 00:31:55,280
And when you want to provide it with even copied text, it doesn't fill up your prompt.

509
00:31:55,280 --> 00:32:00,120
It just gets added as an attachment.

510
00:32:00,120 --> 00:32:05,800
So you can upload PDFs, Word docs, copy paste large chunks of text and they'll basically

511
00:32:05,800 --> 00:32:07,840
be added as a text attachment.

512
00:32:07,840 --> 00:32:11,480
It just makes it really easy to summarize and work with large documents.

513
00:32:11,480 --> 00:32:20,760
And because it's much better at its understanding than when you're working with ChatGPT and

514
00:32:20,760 --> 00:32:24,600
also when you're working with 3.5, but also when you're working with Claude 1, you can

515
00:32:24,600 --> 00:32:28,840
actually trust the outputs much better and you get much more high quality outputs from

516
00:32:28,840 --> 00:32:30,600
it.

517
00:32:30,600 --> 00:32:35,440
Because it's a new model, its knowledge curve is early 2023 versus 2021 for ChatGPT.

518
00:32:35,440 --> 00:32:42,240
It also knows a bit more about the world and it's just well worth a play with if you've

519
00:32:42,240 --> 00:32:46,640
been on the fence about having a play with Claude, now is the time because it's getting

520
00:32:46,640 --> 00:32:50,840
up towards GPT-4 capabilities in some areas.

521
00:32:50,840 --> 00:32:52,800
And it's free.

522
00:32:52,800 --> 00:32:53,800
And it's free.

523
00:32:53,800 --> 00:32:58,640
Yeah, you don't have to pay for all of these additional plugins.

524
00:32:58,640 --> 00:33:05,320
The up-to-date knowledge piece is so welcome.

525
00:33:05,320 --> 00:33:10,080
I was playing around with it earlier this week and it's just nice to be able to ask

526
00:33:10,080 --> 00:33:16,420
it questions about the latest research in certain areas and it knows, it doesn't say

527
00:33:16,420 --> 00:33:19,240
I'm not familiar with it.

528
00:33:19,240 --> 00:33:20,640
That's actually one of the big selling points.

529
00:33:20,640 --> 00:33:24,260
It's been integrated already into Jasper, the AI copywriting tool.

530
00:33:24,260 --> 00:33:28,360
They've actually cited the up-to-date knowledge training as one of the key reasons that they've

531
00:33:28,360 --> 00:33:32,600
integrated it into the models that they now use.

532
00:33:32,600 --> 00:33:34,440
Yeah, it's exciting.

533
00:33:34,440 --> 00:33:38,880
I think it's definitely worth going and having a play with that if you haven't so far.

534
00:33:38,880 --> 00:33:43,880
Although, of course, when a product is free, who's the product, Martin?

535
00:33:43,880 --> 00:33:44,880
You.

536
00:33:44,880 --> 00:33:45,880
That's right.

537
00:33:45,880 --> 00:33:46,880
We're all the products.

538
00:33:46,880 --> 00:33:49,560
So the standard caveats come.

539
00:33:49,560 --> 00:33:54,760
Don't paste in any information that you wouldn't be comfortable going out into the world.

540
00:33:54,760 --> 00:33:55,760
I read somewhere.

541
00:33:55,760 --> 00:34:01,680
I'm not sure I entirely agree with this, but don't put anything into a free to use model

542
00:34:01,680 --> 00:34:07,680
or even chat GPT paid model that you wouldn't be willing to anonymously post on Reddit.

543
00:34:07,680 --> 00:34:11,800
That's the bath I heard for what you should and shouldn't put into these models.

544
00:34:11,800 --> 00:34:16,640
That seems like a pretty high bar to me, but I think especially if you are on the risk

545
00:34:16,640 --> 00:34:21,320
of our side or you're handling information that would be considered commercially sensitive,

546
00:34:21,320 --> 00:34:24,200
it's probably a good bar to set.

547
00:34:24,200 --> 00:34:27,760
Right, more language model news.

548
00:34:27,760 --> 00:34:32,720
Meta has released a commercial version of its open source language model, Lama 2, so

549
00:34:32,720 --> 00:34:36,160
that businesses can build products on top of it.

550
00:34:36,160 --> 00:34:41,640
It's available in three model sizes, ranging from seven billion to 70 billion parameters.

551
00:34:41,640 --> 00:34:48,240
So that is now basically Meta taking what it originally released as open source, but

552
00:34:48,240 --> 00:34:53,960
without any commercial use whatsoever and making it available for companies to use.

553
00:34:53,960 --> 00:34:58,440
The first company I saw jump on this was Perplexity AI, which was a tool of the week that you

554
00:34:58,440 --> 00:35:02,880
introduced us to a few months ago or a few weeks ago, mine.

555
00:35:02,880 --> 00:35:05,320
And it's blazing fast.

556
00:35:05,320 --> 00:35:06,320
Isn't it?

557
00:35:06,320 --> 00:35:07,320
Wow.

558
00:35:07,320 --> 00:35:09,000
It's like real time almost.

559
00:35:09,000 --> 00:35:12,840
And the quality of the outputs in some of my initial blogging tests, I got a blogging

560
00:35:12,840 --> 00:35:19,240
test that I run all new models through to see how nuanced the outline is and how well

561
00:35:19,240 --> 00:35:22,160
crafted the copy is.

562
00:35:22,160 --> 00:35:26,000
It was closer to GPT-4 than I thought it was going to be.

563
00:35:26,000 --> 00:35:31,160
Which is pretty awesome in this particular use case.

564
00:35:31,160 --> 00:35:36,880
There is a note that Meta have put some restrictions on the commercial use of this tool for, I

565
00:35:36,880 --> 00:35:40,640
think it's if you've got more than 700 million users.

566
00:35:40,640 --> 00:35:42,760
Is that sad about right, Martin?

567
00:35:42,760 --> 00:35:43,760
Yeah.

568
00:35:43,760 --> 00:35:49,680
So they basically stopped the likes of Google just being able to use it or anyone like that.

569
00:35:49,680 --> 00:35:54,080
And there's also some weird caveat about, I'm not sure of the exact phrasing, but it's

570
00:35:54,080 --> 00:35:58,720
like if you're planning to use this to solve world hunger, or if you're using the model

571
00:35:58,720 --> 00:36:04,160
for some major world changing event, like there's some restrictions on it.

572
00:36:04,160 --> 00:36:10,760
So I know people from the open source community, like leaders within that field have basically

573
00:36:10,760 --> 00:36:15,360
said, yeah, they said it's open source, but it isn't open open source.

574
00:36:15,360 --> 00:36:22,000
But it's a step in the right direction, certainly much more so than open AI, ironically.

575
00:36:22,000 --> 00:36:23,000
Yeah.

576
00:36:23,000 --> 00:36:26,200
Closed AI, as we're going to start calling them on the podcast.

577
00:36:26,200 --> 00:36:27,600
But yeah, I agree.

578
00:36:27,600 --> 00:36:32,760
I read somewhere that the 700 million users seemed quite like an interesting cutoff that

579
00:36:32,760 --> 00:36:36,760
would be very specific, that would allow you to very easily say, okay, it's not going to

580
00:36:36,760 --> 00:36:38,680
be Google, it's not going to be Amazon, right?

581
00:36:38,680 --> 00:36:42,440
And all these other platforms that are clearly competitors to Meta.

582
00:36:42,440 --> 00:36:47,360
I also heard someone say, yeah, maybe it was also driven by the size of company they thought

583
00:36:47,360 --> 00:36:48,760
they could easily buy.

584
00:36:48,760 --> 00:36:57,520
It's like, up to about 700 million users, we could buy them if we had to, as they grew.

585
00:36:57,520 --> 00:36:59,080
But outside of that would be too difficult.

586
00:36:59,080 --> 00:37:00,840
So I thought that was quite interesting.

587
00:37:00,840 --> 00:37:04,040
I had one quick play with it.

588
00:37:04,040 --> 00:37:08,320
It was on the 7 billion parameter model.

589
00:37:08,320 --> 00:37:12,160
I didn't realize that was the one I'd selected, but it was the first test I did on perplexity.

590
00:37:12,160 --> 00:37:20,280
And I asked it to describe the story of Goldilocks, and it gave me the most weird response.

591
00:37:20,280 --> 00:37:28,520
And it was basically saying that there were troubling themes to do with Goldilocks, and

592
00:37:28,520 --> 00:37:32,200
it couldn't describe Goldilocks because of all of these troubling themes.

593
00:37:32,200 --> 00:37:39,120
And really, we should all move on from the story of Goldilocks because of its troubling

594
00:37:39,120 --> 00:37:40,120
themes.

595
00:37:40,120 --> 00:37:42,360
I thought, how has that come across?

596
00:37:42,360 --> 00:37:46,080
Clearly there's some guardrails that have been put around this.

597
00:37:46,080 --> 00:37:50,720
The prompt was literally describing the story of Goldilocks, and I could never get it to

598
00:37:50,720 --> 00:37:51,720
recreate it.

599
00:37:51,720 --> 00:37:58,160
But it gave me a multi-paragraph response telling me why Goldilocks was bad and the

600
00:37:58,160 --> 00:38:01,320
themes in it were not suitable for basically modern society.

601
00:38:01,320 --> 00:38:05,680
It was as if I'd asked it to write a poem praising Adolf Hitler.

602
00:38:05,680 --> 00:38:11,080
Right, so we've done OpenAI, we've done a bit of meta news.

603
00:38:11,080 --> 00:38:12,480
Let's talk Google for a bit.

604
00:38:12,480 --> 00:38:15,880
There's quite a bit of Google news over the last couple of weeks as well.

605
00:38:15,880 --> 00:38:20,040
So I think first and foremost, there's been a ton of updates to BARD.

606
00:38:20,040 --> 00:38:25,040
So you can now upload images as part of your prompt via an integration with Google Lens.

607
00:38:25,040 --> 00:38:29,200
I saw some people trying to do some interesting stuff with this one, Martin, just this image

608
00:38:29,200 --> 00:38:34,160
one, like uploading receipts and things to see if it could easily pull information out

609
00:38:34,160 --> 00:38:39,840
to help expensing, but it was making stuff up, so it couldn't really be trusted.

610
00:38:39,840 --> 00:38:49,760
I tried it when I was in France and there was an advert of a chicken and there was a

611
00:38:49,760 --> 00:38:54,240
chicken poking his head in the side, looking at the user and another chicken next to it

612
00:38:54,240 --> 00:38:55,600
with an egg coming out of it.

613
00:38:55,600 --> 00:38:57,840
And it said something in French about chat GPT.

614
00:38:57,840 --> 00:39:00,280
And I had zero context for what this advert was.

615
00:39:00,280 --> 00:39:03,060
I was just like, this makes no sense to me.

616
00:39:03,060 --> 00:39:08,640
So I took a photo of it and asked BARD to tell me why it was funny and what was going

617
00:39:08,640 --> 00:39:10,200
on.

618
00:39:10,200 --> 00:39:16,680
And while it did a good job, it read the text in the image and translated that accurately,

619
00:39:16,680 --> 00:39:19,160
but it didn't quite get what was going on in the image.

620
00:39:19,160 --> 00:39:22,320
So it thought the chicken was wearing sunglasses, which it wasn't.

621
00:39:22,320 --> 00:39:30,840
And then it went on and on this big explanation about why it was funny because it was juxtaposing

622
00:39:30,840 --> 00:39:36,120
a chicken laying an egg with a chicken wearing sunglasses and ha ha ha, isn't that funny?

623
00:39:36,120 --> 00:39:38,600
And I just thought, well, I mean, fair play to it.

624
00:39:38,600 --> 00:39:43,560
It did a good job kind of getting an understanding of what was on the image with the text.

625
00:39:43,560 --> 00:39:48,360
But yeah, after that, it all sort of fell down.

626
00:39:48,360 --> 00:39:49,680
Yeah.

627
00:39:49,680 --> 00:39:54,620
It's interesting that there have been tools for a decade or more that could pull in text

628
00:39:54,620 --> 00:39:58,280
information out of printed materials.

629
00:39:58,280 --> 00:40:04,680
It wouldn't appear to be that hard necessarily to add that capability.

630
00:40:04,680 --> 00:40:07,840
But I guess the way these models are trained is very different.

631
00:40:07,840 --> 00:40:11,440
And so it's not an Apple and Pears comparison.

632
00:40:11,440 --> 00:40:15,600
It makes you wonder what's going on in the back end, because as you say, like it's screen

633
00:40:15,600 --> 00:40:16,600
reading.

634
00:40:16,600 --> 00:40:20,560
So we've had OCR for text and images for a long time.

635
00:40:20,560 --> 00:40:24,080
So when you upload an image, does it run that process first?

636
00:40:24,080 --> 00:40:28,440
Are there multiple things that it does in the processing?

637
00:40:28,440 --> 00:40:34,080
Like does Google, does it do a Google lens job of describing the image, extracting the

638
00:40:34,080 --> 00:40:38,520
text separately, turning all of that into a prompt, aligning that with your written

639
00:40:38,520 --> 00:40:42,680
prompt and then kind of spitting out the answer?

640
00:40:42,680 --> 00:40:47,560
I'm really interested to see what's going on behind the scenes.

641
00:40:47,560 --> 00:40:53,760
Yeah, because the compute power that would be required to chain all those different analyses

642
00:40:53,760 --> 00:40:59,640
together, especially if a bunch of them are not needed, right, would be a lot.

643
00:40:59,640 --> 00:41:03,680
That's going to be interesting to see how that plays out over time.

644
00:41:03,680 --> 00:41:08,680
And comes back to this concept of lots of different AI tools that are really good at

645
00:41:08,680 --> 00:41:13,720
a handful of things versus one master AI tool, unless we can figure out how to get the compute

646
00:41:13,720 --> 00:41:15,760
costs and time down.

647
00:41:15,760 --> 00:41:19,920
Because chaining all those different analyses together is going to add a lot there.

648
00:41:19,920 --> 00:41:23,880
There's some other cool stuff that came with some of the updates to Bard.

649
00:41:23,880 --> 00:41:30,800
So since I started playing with Bard, you could use your microphone to record your prompt.

650
00:41:30,800 --> 00:41:35,680
So it's probably been around launch and not, you know, if it came out afterwards, not long

651
00:41:35,680 --> 00:41:36,680
afterwards.

652
00:41:36,680 --> 00:41:39,920
Now Bard can read the responses to you.

653
00:41:39,920 --> 00:41:47,280
So we're drifting much more into a Star Trek, Hello Computer or Jasper style worlds of interacting

654
00:41:47,280 --> 00:41:53,120
with these large language models in a very natural you speak and then it speaks back

655
00:41:53,120 --> 00:41:56,080
to you type way.

656
00:41:56,080 --> 00:41:58,720
Which at some point then how does that work?

657
00:41:58,720 --> 00:42:04,480
Are we going to be able to dig out our Alexa's and our Google Homes and actually make them

658
00:42:04,480 --> 00:42:09,860
work for more than just set a timer for me cooking these eggs and what's the weather

659
00:42:09,860 --> 00:42:11,880
like later on?

660
00:42:11,880 --> 00:42:14,040
It's the next logical step, isn't it?

661
00:42:14,040 --> 00:42:18,120
Particularly with Google Home and I think they've already said that Bard will be integrated

662
00:42:18,120 --> 00:42:21,640
with Google Home.

663
00:42:21,640 --> 00:42:24,720
Amazon's approach to this is going to be interesting.

664
00:42:24,720 --> 00:42:29,480
They are noticeable in their absence of not having a large language model instead offering

665
00:42:29,480 --> 00:42:30,480
a marketplace.

666
00:42:30,480 --> 00:42:36,080
Are they going to partner with the likes of AI21 or are they going to launch their own

667
00:42:36,080 --> 00:42:40,880
language model and integrate it with the smart speaker system whose name I'm not saying right

668
00:42:40,880 --> 00:42:46,400
now because I'm sat next to one of their speakers and it will respond to me when I really don't

669
00:42:46,400 --> 00:42:47,400
want it to.

670
00:42:47,400 --> 00:42:51,880
But yeah, that's the way it's going.

671
00:42:51,880 --> 00:42:57,000
Yeah, I think it would be interesting to see how these are playing out.

672
00:42:57,000 --> 00:43:02,040
I'm using a tool that's better than the one that I built for myself called Audio Pen so

673
00:43:02,040 --> 00:43:06,760
that if I'm walking the dog, I can leave an audio note and it's auto transcribed and summarized

674
00:43:06,760 --> 00:43:09,120
in one of a number of styles.

675
00:43:09,120 --> 00:43:13,360
Personally, I would love to get to the point where I can have a smart speaker, although

676
00:43:13,360 --> 00:43:17,960
I guess the headset I've got on is just as good through the browser and just speak stuff.

677
00:43:17,960 --> 00:43:24,720
I'd love to just talk to Google Mail and then have it draft emails for me in my style that

678
00:43:24,720 --> 00:43:28,080
are more than the ramblings of how I speak, right?

679
00:43:28,080 --> 00:43:34,680
But that can't maintain the essence that is how I want to communicate the words I typically

680
00:43:34,680 --> 00:43:35,840
say.

681
00:43:35,840 --> 00:43:40,420
The use case that I really want to see for this is within the car.

682
00:43:40,420 --> 00:43:45,760
So Android Auto or Apple CarPlay, whatever it's called.

683
00:43:45,760 --> 00:43:47,640
That's the thing that's missing for me.

684
00:43:47,640 --> 00:43:52,360
The fact that there isn't an app so far that I can find on Android Auto where I can record

685
00:43:52,360 --> 00:43:54,440
a voice memo hands free.

686
00:43:54,440 --> 00:43:59,640
If I want to record a voice memo in the car, I have to actually press it on my device,

687
00:43:59,640 --> 00:44:01,720
which is not ideal.

688
00:44:01,720 --> 00:44:09,160
So I'm trying to minimize the use of that if you could just do it with Android Auto

689
00:44:09,160 --> 00:44:12,560
and it's integrated with Bard and you can have that back and forth.

690
00:44:12,560 --> 00:44:16,800
That's a game changer, particularly if you're doing long commutes and lots of driving.

691
00:44:16,800 --> 00:44:17,800
Yeah.

692
00:44:17,800 --> 00:44:20,440
Have a conversation with the car.

693
00:44:20,440 --> 00:44:23,480
But a productive work based one where I can then send the email for you afterwards.

694
00:44:23,480 --> 00:44:24,960
You're like, read that back to me.

695
00:44:24,960 --> 00:44:25,960
Cool.

696
00:44:25,960 --> 00:44:27,560
That's bang on what I want.

697
00:44:27,560 --> 00:44:28,560
Send it off to Bob.

698
00:44:28,560 --> 00:44:32,840
Well, you can do that already with WhatsApp messages in Android Auto, can't you?

699
00:44:32,840 --> 00:44:33,840
Oh, can you?

700
00:44:33,840 --> 00:44:37,720
I don't have Android Auto, so I don't know.

701
00:44:37,720 --> 00:44:38,720
Interesting.

702
00:44:38,720 --> 00:44:39,720
Okay.

703
00:44:39,720 --> 00:44:42,200
What else can you do?

704
00:44:42,200 --> 00:44:45,280
There's a bit of other stuff you can do.

705
00:44:45,280 --> 00:44:49,880
I just want to mention when Google talks to you, it's quite fun to play with it.

706
00:44:49,880 --> 00:44:51,480
It's not as good as Eleven Labs.

707
00:44:51,480 --> 00:44:55,960
So we've talked about on the podcast before Eleven Labs is a company that's trying to

708
00:44:55,960 --> 00:45:02,360
solve for creating synthetic voices that sound like human voices, both in terms of the sound

709
00:45:02,360 --> 00:45:07,840
of a voice, but also the natural pattern of how people speak with pauses and tone changes

710
00:45:07,840 --> 00:45:10,720
and volume changes.

711
00:45:10,720 --> 00:45:14,960
Bard's ability to read things out is not quite as good as that, but it's certainly a far

712
00:45:14,960 --> 00:45:21,300
cry from old school style robotic text to speech readers.

713
00:45:21,300 --> 00:45:24,360
So we're getting there in short.

714
00:45:24,360 --> 00:45:25,360
What else can you do?

715
00:45:25,360 --> 00:45:28,000
You can modify the responses a little bit like you can in Bing.

716
00:45:28,000 --> 00:45:32,640
So you can say, I want the response to be simple or long or short, professional or casual.

717
00:45:32,640 --> 00:45:37,320
So you can change some of that stuff if you're having it create content for you, which I

718
00:45:37,320 --> 00:45:42,120
wouldn't recommend because it's nowhere near as good as Claude or ChatGPT, but you can

719
00:45:42,120 --> 00:45:43,940
if you want.

720
00:45:43,940 --> 00:45:46,180
It's easier to share the responses now via link.

721
00:45:46,180 --> 00:45:49,400
So if you get a response and you think you want to share that with someone, you can just

722
00:45:49,400 --> 00:45:53,160
create a link that you can populate and say an email to send someone.

723
00:45:53,160 --> 00:45:59,400
And Bard is now available in Europe without VPN after addressing privacy concerns.

724
00:45:59,400 --> 00:46:01,520
So lots of updates on Bard.

725
00:46:01,520 --> 00:46:06,240
Our take home is still what we just said, which is not as good as the other tools really.

726
00:46:06,240 --> 00:46:07,520
Got internet access.

727
00:46:07,520 --> 00:46:13,720
That's a massive boon, especially now ChatGPT doesn't unless you use Martin's clever plugin

728
00:46:13,720 --> 00:46:14,880
work around.

729
00:46:14,880 --> 00:46:15,880
Right.

730
00:46:15,880 --> 00:46:18,120
Let's keep the Google news going.

731
00:46:18,120 --> 00:46:23,140
We've got Google launching its AI powered notes app called Notebook LM.

732
00:46:23,140 --> 00:46:27,920
So you'll remember this when Martin told us about all the news that came out of Google's

733
00:46:27,920 --> 00:46:29,880
I think it was Google's developer conference, was it?

734
00:46:29,880 --> 00:46:30,880
Is it IO?

735
00:46:30,880 --> 00:46:31,880
Yeah.

736
00:46:31,880 --> 00:46:37,800
So we knew about this, but the important and interesting thing here is now you can access

737
00:46:37,800 --> 00:46:38,800
it.

738
00:46:38,800 --> 00:46:42,720
It's only available to select users in the US and there is a waitlist you can join.

739
00:46:42,720 --> 00:46:48,480
It is cited by Google themselves as an experimental product designed to use the power and promise

740
00:46:48,480 --> 00:46:49,480
of language models.

741
00:46:49,480 --> 00:46:53,720
That sounds like corporate language paired with your existing content to gain critical

742
00:46:53,720 --> 00:46:55,440
insights faster.

743
00:46:55,440 --> 00:46:57,720
Basically, how does it work?

744
00:46:57,720 --> 00:47:04,000
It is a virtual assistant chat bot style tool that has access to your document library.

745
00:47:04,000 --> 00:47:09,240
It looks at the moment like you have to give it three, four, five documents to like set

746
00:47:09,240 --> 00:47:12,440
the scene that you want to have a conversation about those.

747
00:47:12,440 --> 00:47:20,440
So it might be that you've got an internal memo and a document like a blog post, I don't

748
00:47:20,440 --> 00:47:24,320
know, and some stats and data that someone pulled out and dropped into a Google doc for

749
00:47:24,320 --> 00:47:25,320
you.

750
00:47:25,320 --> 00:47:27,520
You can query those all in one go.

751
00:47:27,520 --> 00:47:32,680
What this seems like to me is the early versions of what co-pilots are going to do once they

752
00:47:32,680 --> 00:47:38,240
get baked into workspace and also into Office 365.

753
00:47:38,240 --> 00:47:42,920
So if you can get access to that, it's worth getting on the waitlist and having a little

754
00:47:42,920 --> 00:47:44,360
bit of a play.

755
00:47:44,360 --> 00:47:48,720
I'm struggling to see the use case for this that isn't quickly asserted by those co-pilots

756
00:47:48,720 --> 00:47:53,920
Martin now that I know how it's going to work.

757
00:47:53,920 --> 00:47:54,920
Martin agrees.

758
00:47:54,920 --> 00:47:55,920
He nods excitedly.

759
00:47:55,920 --> 00:48:01,400
He says, move on to the next story about Genesis.

760
00:48:01,400 --> 00:48:02,400
Can do.

761
00:48:02,400 --> 00:48:06,920
The band could talk for hours about the band, but it isn't the band on this occasion.

762
00:48:06,920 --> 00:48:10,880
It's Google testing an AI tool called Genesis that can write news articles.

763
00:48:10,880 --> 00:48:15,400
Google's been about town pitching this to publications like the New York Times, Washington

764
00:48:15,400 --> 00:48:17,920
Post, et cetera, et cetera.

765
00:48:17,920 --> 00:48:22,160
They state very clearly that the AI enabled tools are not meant to replace journalists,

766
00:48:22,160 --> 00:48:26,960
but rather act as co-pilots and that Google is only in the early stages of exploring the

767
00:48:26,960 --> 00:48:31,760
tool, which it sees as being an assistant that helps with things like headlines and

768
00:48:31,760 --> 00:48:36,480
changing copyright writing styles.

769
00:48:36,480 --> 00:48:39,840
As reported by The Guardian, there's a Guardian article on this, two executives at the New

770
00:48:39,840 --> 00:48:44,040
York Times who saw the pitch said, it seemed to take for granted the effort that went into

771
00:48:44,040 --> 00:48:47,960
producing accurate and artful news stories.

772
00:48:47,960 --> 00:48:52,240
I think our take on this Martin is that major parts of journalism are still well out of

773
00:48:52,240 --> 00:48:56,600
reach for current AIs, not to mention the fact checking that we've been talking a lot

774
00:48:56,600 --> 00:49:01,440
about today and over the last couple of months.

775
00:49:01,440 --> 00:49:06,920
Just in play by Google, like going into senior leadership, try and tell me that that is not

776
00:49:06,920 --> 00:49:09,600
an argument to cut writing stuff.

777
00:49:09,600 --> 00:49:12,240
Ballsy, isn't it?

778
00:49:12,240 --> 00:49:13,240
Absolutely.

779
00:49:13,240 --> 00:49:20,280
Particularly in the landscape where we're seeing the Hollywood actors strikes and the

780
00:49:20,280 --> 00:49:26,800
writers strikes, all of which is concerned with how AI is going to be used within creative

781
00:49:26,800 --> 00:49:34,200
production workflows, actors being able to record a thing, and then the AI edits what

782
00:49:34,200 --> 00:49:38,520
they say, they don't have to be coming back in for reshoots, meaning they don't get paid

783
00:49:38,520 --> 00:49:39,520
as much.

784
00:49:39,520 --> 00:49:41,880
All of these kinds of things.

785
00:49:41,880 --> 00:49:47,160
If Google is pitching, hey, your sub editors, you can get rid of them.

786
00:49:47,160 --> 00:49:48,160
That's a strong move.

787
00:49:48,160 --> 00:49:53,640
It sounds like the top brass are not overly convinced though.

788
00:49:53,640 --> 00:49:54,640
Not publicly.

789
00:49:54,640 --> 00:49:58,080
Maybe they're like, Google, this sounds amazing.

790
00:49:58,080 --> 00:50:01,680
I am going to have to come out in the press and say it sounds like Tosh and that it will

791
00:50:01,680 --> 00:50:03,720
never work and that we'd never do it.

792
00:50:03,720 --> 00:50:06,000
But yeah, do tell me more.

793
00:50:06,000 --> 00:50:07,600
It's difficult to say.

794
00:50:07,600 --> 00:50:11,880
On the writers and the Hollywood strike, I don't know how much credence there is in this

795
00:50:11,880 --> 00:50:18,520
story, but it's been floated that some of the major studios are trying to get supporting

796
00:50:18,520 --> 00:50:24,320
actors, so background actors, to come in for a day's worth of work and get scanned as a

797
00:50:24,320 --> 00:50:26,560
basically extras.

798
00:50:26,560 --> 00:50:33,320
Then use them across multiple productions across their suite of whatever they're producing,

799
00:50:33,320 --> 00:50:37,640
but only have to pay the people once because they get the scan of them and then away they

800
00:50:37,640 --> 00:50:38,640
go.

801
00:50:38,640 --> 00:50:44,240
Oh, well, you can totally buy that, can't you?

802
00:50:44,240 --> 00:50:51,920
I was working with a company that does virtual production, an interesting area of TV and

803
00:50:51,920 --> 00:50:58,440
movie production where you have large LED screens, so you'll have a real scene with

804
00:50:58,440 --> 00:51:04,560
real actors in the foreground, but all in the background is just an LED projection.

805
00:51:04,560 --> 00:51:10,320
What you can do with that is if you use a slightly out of focus background scene, you

806
00:51:10,320 --> 00:51:16,040
can just have a crowd scene or exactly like I say, background actors.

807
00:51:16,040 --> 00:51:21,720
If they're slightly out of focus, you just do it all using Unreal Engine.

808
00:51:21,720 --> 00:51:25,960
It looks good enough when it goes onto the big screen to be real.

809
00:51:25,960 --> 00:51:27,240
Already that's not AI generated.

810
00:51:27,240 --> 00:51:34,080
We're already seeing the need for real people be taken out of production at some extent,

811
00:51:34,080 --> 00:51:38,320
so it makes perfect sense to me that they would try to do that as well.

812
00:51:38,320 --> 00:51:39,320
Agreed.

813
00:51:39,320 --> 00:51:45,720
I think it's a bit of a tangent, but it really, really puts a number of creative professionals

814
00:51:45,720 --> 00:51:52,640
when you look at it in this instance between a rock and a hard place because if synthetic,

815
00:51:52,640 --> 00:51:57,560
you talk about blurring those synthetic people created in Unreal Engine, which is a gaming

816
00:51:57,560 --> 00:52:02,080
engine, traditionally a gaming engine, getting powerful enough to potentially be used in

817
00:52:02,080 --> 00:52:08,440
movies, but if you can create synthetic people, it's like, well, come in and get a day's

818
00:52:08,440 --> 00:52:13,520
worth of work and let me scan you or don't and I'll just have synthetic people in the

819
00:52:13,520 --> 00:52:14,520
background.

820
00:52:14,520 --> 00:52:22,960
So, Krums, that is an industry conundrum that is probably going to be an interesting trend

821
00:52:22,960 --> 00:52:30,880
setter for how other areas of the economy have to deal with the be augmented or be replaced.

822
00:52:30,880 --> 00:52:36,160
There's going to be a bunch of areas of the economy, probably many more areas of the economy

823
00:52:36,160 --> 00:52:40,720
where you can't do this, but in those areas where you can, it doesn't give the people

824
00:52:40,720 --> 00:52:43,120
much bargaining power, does it?

825
00:52:43,120 --> 00:52:44,160
No.

826
00:52:44,160 --> 00:52:51,400
We should all just have careers as refuge collectors and refuse collectors even.

827
00:52:51,400 --> 00:52:56,480
You've got to be a bin man, can't AI that way.

828
00:52:56,480 --> 00:53:00,120
Well unless Optimus and these other robots improve in performance significantly over

829
00:53:00,120 --> 00:53:07,440
the next five or so years, but yeah, I definitely think if I was like 18 now, I think I'd be

830
00:53:07,440 --> 00:53:13,160
like, right, electrician, plumbing, roofing, the hard to automate.

831
00:53:13,160 --> 00:53:14,160
Yeah.

832
00:53:14,160 --> 00:53:17,320
Anyway, we digress.

833
00:53:17,320 --> 00:53:21,120
Let's talk HubSpot for a bit.

834
00:53:21,120 --> 00:53:26,880
HubSpot's ChatSpot, say that after you've had a few drinks at a wedding, Martin, now

835
00:53:26,880 --> 00:53:30,080
learns your writing style for blogs and emails.

836
00:53:30,080 --> 00:53:34,440
So the launch email on this claimed that you could train your ChatSpot account on up to

837
00:53:34,440 --> 00:53:39,440
5,000 words of text from your blog or previous email campaigns and it will use this to get

838
00:53:39,440 --> 00:53:44,080
closer to your brand style when creating new blog or email content.

839
00:53:44,080 --> 00:53:49,480
However, when you played with this, Martin, you didn't quite get that as your experience,

840
00:53:49,480 --> 00:53:50,680
did you?

841
00:53:50,680 --> 00:53:51,680
Not even close.

842
00:53:51,680 --> 00:53:57,280
So the emails where they launched it said 5,000 words and it's 5,000 characters.

843
00:53:57,280 --> 00:54:01,560
And that is a dramatic difference in terms of training input.

844
00:54:01,560 --> 00:54:06,240
So it gives you around 700 to 1200 words that you can input.

845
00:54:06,240 --> 00:54:12,840
So that's basically one blog post, realistically, that you can train it on.

846
00:54:12,840 --> 00:54:15,840
So I was a bit surprised when I came to actually use it.

847
00:54:15,840 --> 00:54:19,800
It's 5,000 characters for email as well.

848
00:54:19,800 --> 00:54:27,080
That's quite a few emails, so you can actually put in several examples.

849
00:54:27,080 --> 00:54:32,680
The blog version needs a bit more work.

850
00:54:32,680 --> 00:54:41,800
Yeah, actually, when you correct the error, it's much more in the context windows of ChatGPT

851
00:54:41,800 --> 00:54:44,560
and the APIs underpinning that, right?

852
00:54:44,560 --> 00:54:52,720
5,000 characters is probably, I don't know, 8,000 tokens, 16,000 tokens sounds a bit more

853
00:54:52,720 --> 00:54:54,800
realistic perhaps.

854
00:54:54,800 --> 00:54:59,360
But won't be that long before we can maybe, but not this tall yet.

855
00:54:59,360 --> 00:55:02,560
I thought what it does when it analyzes, it was quite interesting.

856
00:55:02,560 --> 00:55:09,520
It gives you a paragraph giving you quite a detailed description of the tone of voice,

857
00:55:09,520 --> 00:55:14,440
which I'm assuming what it then does in the back end is injects that into a prompt when

858
00:55:14,440 --> 00:55:17,360
you then ask it to write a blog.

859
00:55:17,360 --> 00:55:19,160
I think you're absolutely right.

860
00:55:19,160 --> 00:55:23,960
I use a Chrome plugin called Prompt Manager.

861
00:55:23,960 --> 00:55:28,960
And in one of the newsletters I'm subscribed to, I saw an example where you paste a load

862
00:55:28,960 --> 00:55:35,080
of text that you've written into ChatGPT with the prompt that you want it to create you.

863
00:55:35,080 --> 00:55:36,400
What does it call it?

864
00:55:36,400 --> 00:55:43,520
Some sort of voice paragraph or something, something like that, where it basically boils

865
00:55:43,520 --> 00:55:49,600
down the one 2,000 words that you give it to a statement designed to mimic your style,

866
00:55:49,600 --> 00:55:54,480
but in a reductionist sense that you can then use every time you want to write something.

867
00:55:54,480 --> 00:55:57,480
So yeah, that would be similar to that, but it's doing it all in the back end.

868
00:55:57,480 --> 00:55:59,280
So you don't have to worry about including that.

869
00:55:59,280 --> 00:56:06,480
Of course, if you've got access to the new custom tool from, it's been so much that we've

870
00:56:06,480 --> 00:56:07,680
covered today, what's it called?

871
00:56:07,680 --> 00:56:09,880
Custom instructions.

872
00:56:09,880 --> 00:56:12,320
You could do that in your custom instructions as well now, couldn't you?

873
00:56:12,320 --> 00:56:16,400
So that would be quite a good use case.

874
00:56:16,400 --> 00:56:18,080
Right.

875
00:56:18,080 --> 00:56:27,000
Let's talk Microsoft's OpenAI features coming into Office AI.

876
00:56:27,000 --> 00:56:33,600
It's been reported that Microsoft are going to charge $30 per user for you to use AI in

877
00:56:33,600 --> 00:56:36,720
Office 365 products.

878
00:56:36,720 --> 00:56:38,200
I've heard a bit of grumbling about this.

879
00:56:38,200 --> 00:56:39,200
Oh, go on, mate.

880
00:56:39,200 --> 00:56:43,840
Because it is a $30 per user per month, right?

881
00:56:43,840 --> 00:56:44,840
Yes.

882
00:56:44,840 --> 00:56:51,400
Which is quite, that is more than I was expecting it to be given that, you know, a standard,

883
00:56:51,400 --> 00:56:59,160
let's take home office user is about, what, 70, 80 quid a year.

884
00:56:59,160 --> 00:57:04,040
The enterprise licenses are not much more than that, are they per person per year?

885
00:57:04,040 --> 00:57:06,200
I don't know.

886
00:57:06,200 --> 00:57:09,720
I think we pay about $10 per user at Biostrata.

887
00:57:09,720 --> 00:57:10,720
Yeah.

888
00:57:10,720 --> 00:57:18,000
So going up from 10 to 40, you know, that's a substantial shift.

889
00:57:18,000 --> 00:57:21,520
Well, because it's an interesting one.

890
00:57:21,520 --> 00:57:23,760
And I saw a lot of grumbles online.

891
00:57:23,760 --> 00:57:29,160
Like the main consensus here was like, that's a bit steep.

892
00:57:29,160 --> 00:57:33,560
There are a bunch of tools floating around the internet right now that are $20 per user,

893
00:57:33,560 --> 00:57:38,760
$30 per user, $50 per user, and they're not going to do all the things that copilot is

894
00:57:38,760 --> 00:57:39,760
going to do.

895
00:57:39,760 --> 00:57:42,000
So that's the first thing I would say.

896
00:57:42,000 --> 00:57:48,720
The second thing is just trying to think about how much time this might save someone.

897
00:57:48,720 --> 00:57:52,360
Because I do appreciate that it seems like a lot.

898
00:57:52,360 --> 00:57:56,280
And if you're a large enterprise and you've suddenly got to pay $30 per user per month

899
00:57:56,280 --> 00:57:59,280
for all your folks, that's going to cost you a lot.

900
00:57:59,280 --> 00:58:06,120
And I saw Microsoft's share price responded in kind because people were projecting out

901
00:58:06,120 --> 00:58:09,680
how much extra revenue this could drive per year for Microsoft.

902
00:58:09,680 --> 00:58:14,680
But if it saves an employee five to 10 hours a month, and your average employee is probably

903
00:58:14,680 --> 00:58:20,840
doing 120 to 140 valuable hours per month if they're working like a normal nine to five

904
00:58:20,840 --> 00:58:23,620
job trying to factor in holidays and stuff like that.

905
00:58:23,620 --> 00:58:34,400
So that would be, you know, less than what 10% of their time being saved, right?

906
00:58:34,400 --> 00:58:42,680
If you're paying your employees $10 an hour, you only need to save three hours a month

907
00:58:42,680 --> 00:58:43,680
to break even.

908
00:58:43,680 --> 00:58:49,160
If you're saving 10 hours a month, then you're getting a return on investment.

909
00:58:49,160 --> 00:58:54,880
So ultimately, I think if you look at it, it'll end up potentially being quite cheap

910
00:58:54,880 --> 00:58:56,640
compared to the value it unlocks.

911
00:58:56,640 --> 00:59:02,040
And I'm absolutely certain this is where Microsoft are thinking and starting to solve some of

912
00:59:02,040 --> 00:59:06,400
the conundrum that no one else has looked to solve for yet, which is how do we monetize

913
00:59:06,400 --> 00:59:09,240
the productivity gains offered by AI?

914
00:59:09,240 --> 00:59:12,280
And I think this is the first attempt at that.

915
00:59:12,280 --> 00:59:19,440
But, and there's a big but, in order for this to work and to realize that investment that

916
00:59:19,440 --> 00:59:23,000
we just talked about, that return on investment, companies are going to need to train their

917
00:59:23,000 --> 00:59:29,880
staff to ensure that they and the organization can get the full benefits of using the systems.

918
00:59:29,880 --> 00:59:33,080
And they're going to have to find ways to encourage adoption because I think adoption

919
00:59:33,080 --> 00:59:36,920
is going to be mixed, especially if we look at the early adopters of Chatchie BT potentially

920
00:59:36,920 --> 00:59:40,320
using it less now in June than they were before.

921
00:59:40,320 --> 00:59:43,240
And it was actually starting to go down.

922
00:59:43,240 --> 00:59:46,560
So $30 per month per user seems expensive.

923
00:59:46,560 --> 00:59:52,900
I think it's actually cheap, but only if you train your team on how to maximize the benefits

924
00:59:52,900 --> 00:59:58,400
of the tool and you can overcome the inertia of actually getting them to use it and change

925
00:59:58,400 --> 01:00:02,040
the ways they work today.

926
01:00:02,040 --> 01:00:05,980
Now it's currently being tested by a select group of beta test companies and the wider

927
01:00:05,980 --> 01:00:09,600
rollout date is still not announced.

928
01:00:09,600 --> 01:00:14,160
This is a pain in the bottom for a lot of organizations as far as I'm concerned, who

929
01:00:14,160 --> 01:00:19,360
are probably sat on the fence around, should I give people to pay Jack GPT account?

930
01:00:19,360 --> 01:00:22,880
Should I buy Jasper or writer or any of these other tools?

931
01:00:22,880 --> 01:00:25,640
Which tools can I trust with my data or not?

932
01:00:25,640 --> 01:00:30,040
I suspect a lot of us are just going, hurry up and sort this out for us, Microsoft, so

933
01:00:30,040 --> 01:00:34,400
that we can just work in an ecosystem that we know well, that's passed all our security

934
01:00:34,400 --> 01:00:37,360
checks, that we trust you with our data already.

935
01:00:37,360 --> 01:00:40,200
So Krams, we may as well trust you with this as well.

936
01:00:40,200 --> 01:00:42,400
So we can just roll this out to our employees.

937
01:00:42,400 --> 01:00:46,560
Because I think a lot of businesses are in that gray at the moment where they need an

938
01:00:46,560 --> 01:00:51,720
enterprise level provider to step up and meet their checklist of requirements.

939
01:00:51,720 --> 01:00:53,560
And I just don't think there is one yet.

940
01:00:53,560 --> 01:00:56,080
Oh no, absolutely there isn't.

941
01:00:56,080 --> 01:01:00,640
And we're not heard anything from Google in this domain either.

942
01:01:00,640 --> 01:01:02,600
So we wait patiently.

943
01:01:02,600 --> 01:01:12,040
Well, we've got notebook, right, which is definitely a very Diet Coke version of this.

944
01:01:12,040 --> 01:01:17,720
I do know that a bunch of people have got access to the Google workspace equivalent,

945
01:01:17,720 --> 01:01:20,400
which basically is a little bit like the one in HubSpot.

946
01:01:20,400 --> 01:01:24,120
It gives you a little button that you can click when you select text and you can rewrite

947
01:01:24,120 --> 01:01:26,880
it, expand it and those types of things.

948
01:01:26,880 --> 01:01:33,520
But if it's based on Google's large language models, which are arguably the worst so far,

949
01:01:33,520 --> 01:01:36,360
I'm not that excited for it.

950
01:01:36,360 --> 01:01:41,360
Whereas Microsoft is powered by GPT-4.

951
01:01:41,360 --> 01:01:46,160
We should also say when you charge $30 a month and everybody starts bashing these, we're

952
01:01:46,160 --> 01:01:52,800
going to have to go off world to find the materials we need to create all the GPUs.

953
01:01:52,800 --> 01:01:57,560
We're going to have to trick some asteroids into circulating the earth and turn them into

954
01:01:57,560 --> 01:01:58,560
like cloud.

955
01:01:58,560 --> 01:02:04,880
I mean, that's not cloud, that's stratosphere and above.

956
01:02:04,880 --> 01:02:08,680
That's a different issue when everybody's using these tools all the time.

957
01:02:08,680 --> 01:02:12,480
The amount of compute power and energy could get kind of insane.

958
01:02:12,480 --> 01:02:16,720
But yeah, hurry up, Microsoft, I think there's a bunch of us that just need you to make our

959
01:02:16,720 --> 01:02:21,160
product selection easy here and just give it to us.

960
01:02:21,160 --> 01:02:26,800
Right, last bit about large language models and assistance.

961
01:02:26,800 --> 01:02:32,320
Inflection AI, who we've talked about before, the developers of Pi have developed an $880

962
01:02:32,320 --> 01:02:38,180
million AI supercomputer, even when GPUs are currently scarce.

963
01:02:38,180 --> 01:02:41,320
They do have a major investor and that major investor is Nvidia.

964
01:02:41,320 --> 01:02:45,800
So that might explain how they got hold of those.

965
01:02:45,800 --> 01:02:50,560
They've raised $1.5 billion so far and have a $4 billion valuation.

966
01:02:50,560 --> 01:02:55,840
So we talked about them on the last podcast and how the tool just all it wants to do is

967
01:02:55,840 --> 01:03:00,600
ask you questions, but definitely we all need to keep an eye on inflection.

968
01:03:00,600 --> 01:03:02,320
And I was having some fun with Pi.

969
01:03:02,320 --> 01:03:07,760
I just tripled my productivity with Pi, Martin, because I realized that if you use Pi in WhatsApp,

970
01:03:07,760 --> 01:03:16,360
you can leave Pi a voice note, it auto transcribes it and then writes back to you in text.

971
01:03:16,360 --> 01:03:21,360
So I now wander around my house, leaving my thoughts as rambling voice notes that then

972
01:03:21,360 --> 01:03:24,000
Pi has to untangle and then come back with a response.

973
01:03:24,000 --> 01:03:26,920
And I've found that excellent tool for brainstorming.

974
01:03:26,920 --> 01:03:28,920
It's kind of cool.

975
01:03:28,920 --> 01:03:31,920
If you haven't played with Pi yet, get it on the WhatsApps.

976
01:03:31,920 --> 01:03:32,920
It's free.

977
01:03:32,920 --> 01:03:33,920
It's not free.

978
01:03:33,920 --> 01:03:34,920
You're the product, but it's free to use.

979
01:03:34,920 --> 01:03:37,280
You don't have to pay and it's fun to play with.

980
01:03:37,280 --> 01:03:38,280
I'm into that.

981
01:03:38,280 --> 01:03:42,360
That is a use case I'm going to go away and do immediately after this recording.

982
01:03:42,360 --> 01:03:45,640
Do you know what the best thing is?

983
01:03:45,640 --> 01:03:47,040
I had no idea it would work.

984
01:03:47,040 --> 01:03:52,280
I left the voice note and went, Pi, I don't suppose you can understand this.

985
01:03:52,280 --> 01:03:56,520
Back came the text a couple of seconds later going, yeah, yeah, yeah, no worries, pal.

986
01:03:56,520 --> 01:03:58,840
What do you want to talk about?

987
01:03:58,840 --> 01:04:01,720
Again, because all it does is ask questions.

988
01:04:01,720 --> 01:04:02,720
Don't leave me.

989
01:04:02,720 --> 01:04:03,720
It said, don't leave me.

990
01:04:03,720 --> 01:04:05,200
If you don't speak to me, I don't have any value.

991
01:04:05,200 --> 01:04:07,560
Answer my next question.

992
01:04:07,560 --> 01:04:12,920
I actually think that voice note feature really changes how you interact with it a lot.

993
01:04:12,920 --> 01:04:15,920
Makes me use it more, definitely.

994
01:04:15,920 --> 01:04:17,280
Right.

995
01:04:17,280 --> 01:04:18,280
Everybody take a deep breath.

996
01:04:18,280 --> 01:04:21,160
We're going to move into image generation briefly now.

997
01:04:21,160 --> 01:04:26,240
ClipDrop, which is the image creation manipulation tool from Stability AI, which we love on the

998
01:04:26,240 --> 01:04:30,960
podcast because it's very cheap, like £5 per user per month and got a load of tools

999
01:04:30,960 --> 01:04:31,960
in it.

1000
01:04:31,960 --> 01:04:32,960
They've got a new tool.

1001
01:04:32,960 --> 01:04:33,960
It's called Stable Doodle.

1002
01:04:33,960 --> 01:04:35,600
And it's really kind of interesting.

1003
01:04:35,600 --> 01:04:39,720
You create a sketch of something and then it's used as part of the prompt to drive image

1004
01:04:39,720 --> 01:04:40,720
generation.

1005
01:04:40,720 --> 01:04:44,680
I've only done a few tests with it, but it does give you a hell of a lot more control

1006
01:04:44,680 --> 01:04:49,320
over the image that you get, which of course, when you're trying to learn the art of prompting

1007
01:04:49,320 --> 01:04:54,220
mid-journey to get beautiful images, but have some limited control over what it looks like,

1008
01:04:54,220 --> 01:04:56,900
this is actually a very interesting feature.

1009
01:04:56,900 --> 01:04:58,880
You managed to have a play with this yet, Ian?

1010
01:04:58,880 --> 01:05:00,680
I have not.

1011
01:05:00,680 --> 01:05:02,080
All right.

1012
01:05:02,080 --> 01:05:03,160
Well, get on that.

1013
01:05:03,160 --> 01:05:04,160
I think you'll enjoy it.

1014
01:05:04,160 --> 01:05:05,160
What's the problem with holidays?

1015
01:05:05,160 --> 01:05:06,160
Yeah, there's been...

1016
01:05:06,160 --> 01:05:11,040
I was seeing these updates and just trying to carve away a few minutes from the toddler

1017
01:05:11,040 --> 01:05:14,600
just to play with something was proving very difficult.

1018
01:05:14,600 --> 01:05:21,760
I have been using mid-journey for everything though, just as a...

1019
01:05:21,760 --> 01:05:25,760
I've been using mid-journey for all my slide decks I've been putting together recently,

1020
01:05:25,760 --> 01:05:29,320
just because I can get some really good images.

1021
01:05:29,320 --> 01:05:36,440
And I've got a prompt writer saved in ChatGPT, so I've got a conversation where I've given

1022
01:05:36,440 --> 01:05:42,120
it all of the variables and given it loads of example mid-journey prompts and then ChatGPT.

1023
01:05:42,120 --> 01:05:48,560
I just give it a concept like Google eating your lunch was one that I gave it.

1024
01:05:48,560 --> 01:05:52,560
And it gave me this brilliant prompt that I put into mid-journey and mid-journey gave

1025
01:05:52,560 --> 01:05:57,240
me this fantastic illustration that absolutely nailed it.

1026
01:05:57,240 --> 01:06:00,400
I need you to share that workflow with me after this call.

1027
01:06:00,400 --> 01:06:05,600
And maybe if the listeners are nice to us and they give us some nice feedback online

1028
01:06:05,600 --> 01:06:09,920
on the Twitters and the LinkedIn, maybe we'll share it with them as well.

1029
01:06:09,920 --> 01:06:13,440
Speaking of mid-journey, we talked a little bit about the zoom out feature of mid-journey,

1030
01:06:13,440 --> 01:06:15,520
which is pretty cool on a previous episode.

1031
01:06:15,520 --> 01:06:17,520
Now you can pan, which makes sense.

1032
01:06:17,520 --> 01:06:22,680
It's kind of the same engine zooming out really, creating contextually relevant additional

1033
01:06:22,680 --> 01:06:24,840
image sections around an existing image.

1034
01:06:24,840 --> 01:06:28,320
Now you can pan so you can go left, you can go right, you can go up and down, and you

1035
01:06:28,320 --> 01:06:30,680
can pan multiple times on the same image.

1036
01:06:30,680 --> 01:06:34,840
And if you've got remix mode turned on, you can even change the prompt as you extend the

1037
01:06:34,840 --> 01:06:35,840
image.

1038
01:06:35,840 --> 01:06:39,160
So as you're extending it, you can give it some guidance on what that extension should

1039
01:06:39,160 --> 01:06:40,160
include.

1040
01:06:40,160 --> 01:06:41,400
There are some limitations.

1041
01:06:41,400 --> 01:06:43,880
The resolution decreases when zooming out.

1042
01:06:43,880 --> 01:06:48,560
So that is an issue when you're using some of these tools.

1043
01:06:48,560 --> 01:06:54,080
There is also an inability to pan both horizontally and vertically at the same time.

1044
01:06:54,080 --> 01:06:55,600
You can go left or you can go up.

1045
01:06:55,600 --> 01:06:57,840
You can't go both.

1046
01:06:57,840 --> 01:07:03,400
And I think a lot of people are experiencing the possibility for repetition when panning,

1047
01:07:03,400 --> 01:07:04,880
which makes sense.

1048
01:07:04,880 --> 01:07:09,000
But of course, it's going to create an image that maybe isn't as usable as you were hoping

1049
01:07:09,000 --> 01:07:10,000
for.

1050
01:07:10,000 --> 01:07:12,520
There's more image generation news.

1051
01:07:12,520 --> 01:07:18,140
So Meta has a new image generation model called Chameleon that creates images from text and

1052
01:07:18,140 --> 01:07:22,140
can create text from images, so kind of captioning.

1053
01:07:22,140 --> 01:07:26,520
The model is super efficient and requires five times less compute power and a smaller

1054
01:07:26,520 --> 01:07:29,080
training data set than previous models.

1055
01:07:29,080 --> 01:07:32,200
I think one of the most interesting things about it is it can understand instructions

1056
01:07:32,200 --> 01:07:35,040
to edit existing images.

1057
01:07:35,040 --> 01:07:42,440
So you can edit or remix images and give specific instructions based on objects with specific

1058
01:07:42,440 --> 01:07:43,800
coordinates.

1059
01:07:43,800 --> 01:07:50,960
So you can ask the system to place objects in specific locations in an image, and you

1060
01:07:50,960 --> 01:07:55,000
can edit photos with very specific prompts like change the color.

1061
01:07:55,000 --> 01:07:57,640
Now nobody's got access to this as far as I know.

1062
01:07:57,640 --> 01:07:59,240
I certainly don't.

1063
01:07:59,240 --> 01:08:05,400
So how well it is geared up to do this type of work isn't clear.

1064
01:08:05,400 --> 01:08:08,840
But one of the interesting things about it is we've talked on the podcast about some

1065
01:08:08,840 --> 01:08:15,920
of the limitations of existing tools that basically you can create great images, but

1066
01:08:15,920 --> 01:08:20,020
then you've got to push them into Photoshop if you want to do anything with them.

1067
01:08:20,020 --> 01:08:25,020
This will bring an extension of capabilities eventually one assumes that you can edit images

1068
01:08:25,020 --> 01:08:28,680
through natural language prompts instead.

1069
01:08:28,680 --> 01:08:32,960
Meta's not announced whether or not it will plan to release Chameleon.

1070
01:08:32,960 --> 01:08:38,320
It's certainly not released yet, but we're starting to see ever more innovation in the

1071
01:08:38,320 --> 01:08:42,520
generative image space as well as the generative text space.

1072
01:08:42,520 --> 01:08:48,960
You can well imagine that being integrated into their advertising suite, can't you?

1073
01:08:48,960 --> 01:08:52,680
Then advertisers, you might search for a stock image.

1074
01:08:52,680 --> 01:08:57,800
It will pull in a stock image and then you can just ask it to change in a certain way,

1075
01:08:57,800 --> 01:09:01,760
changing the color of someone's top or changing the scene to a wintery scene or something

1076
01:09:01,760 --> 01:09:02,760
like that.

1077
01:09:02,760 --> 01:09:07,800
If that can be fully integrated into the advertising workflow, I think marketers have got an interesting

1078
01:09:07,800 --> 01:09:09,840
new tool to play with there.

1079
01:09:09,840 --> 01:09:13,360
That's definitely where I would expect Meta to go.

1080
01:09:13,360 --> 01:09:15,160
I completely agree.

1081
01:09:15,160 --> 01:09:16,480
Right.

1082
01:09:16,480 --> 01:09:19,920
We've been speaking for quite a long time now, Mart, and I think we'll give the listeners

1083
01:09:19,920 --> 01:09:22,000
a little break.

1084
01:09:22,000 --> 01:09:24,480
What we'll do, there's still some more news to go.

1085
01:09:24,480 --> 01:09:28,880
Because we've not been around for a couple of weeks, we'll hold onto a little bit of

1086
01:09:28,880 --> 01:09:35,300
this news and we'll summarize it in Friday's podcast, which we'll be back doing our weekly

1087
01:09:35,300 --> 01:09:37,360
podcast again from here on out.

1088
01:09:37,360 --> 01:09:41,080
Obviously, Martin's going to be in the US doing some shiny exciting stuff, so we might

1089
01:09:41,080 --> 01:09:45,840
have to do a little bit of jiggery-pokery on our schedules, but we'll be back weekly

1090
01:09:45,840 --> 01:09:52,360
from here sharing more amazing marketing AI news with you.

1091
01:09:52,360 --> 01:09:53,360
Sounds great.

1092
01:09:53,360 --> 01:09:55,440
It's been a pleasure to be back.

1093
01:09:55,440 --> 01:09:56,640
I've enjoyed chatting with you.

1094
01:09:56,640 --> 01:10:00,120
I love the fact there was so much news we couldn't squeeze it into one podcast and we'll

1095
01:10:00,120 --> 01:10:02,680
have to somehow catch up in the next one.

1096
01:10:02,680 --> 01:10:03,680
Yeah.

1097
01:10:03,680 --> 01:10:05,720
And I'm just looking at how much there is left to go.

1098
01:10:05,720 --> 01:10:07,760
I think that's going to run long as well.

1099
01:10:07,760 --> 01:10:10,640
Listeners, thank you for sticking with us.

1100
01:10:10,640 --> 01:10:12,520
Yeah, we appreciate it.

1101
01:10:12,520 --> 01:10:13,800
Everybody enjoy your weeks.

1102
01:10:13,800 --> 01:10:18,560
Thanks for your time today and look forward to the next episode of Artificially Intelligent

1103
01:10:18,560 --> 01:10:19,560
Marketing.

1104
01:10:19,560 --> 01:10:20,560
Bye, Martin.

1105
01:10:20,560 --> 01:10:21,560
Bye.

1106
01:10:21,560 --> 01:10:25,480
Thank you for listening to Artificially Intelligent Marketing.

1107
01:10:25,480 --> 01:10:31,560
To stay on top of the latest trends, tips and tools in the world of marketing AI, be

1108
01:10:31,560 --> 01:10:33,300
sure to subscribe.

1109
01:10:33,300 --> 01:10:44,400
We look forward to seeing you again next week.