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 --> 00:00:18,520
results from your marketing efforts.

4
00:00:18,520 --> 00:00:22,280
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:22,280 --> 00:00:24,620
Hello everybody.

6
00:00:24,620 --> 00:00:28,240
Welcome to Artificially Intelligent Marketing episode 44.

7
00:00:28,240 --> 00:00:30,960
And you were all thinking, where have the guys gone?

8
00:00:30,960 --> 00:00:32,920
I haven't heard from them for ages.

9
00:00:32,920 --> 00:00:37,240
And we are sorry, but we are here today because it's been a fairly big week in the world of

10
00:00:37,240 --> 00:00:38,240
AI.

11
00:00:38,240 --> 00:00:45,160
And so we thought what better time to bring our podcast back alive than this week.

12
00:00:45,160 --> 00:00:51,040
So it's me, Paul Avery here, co-host as usual with my fantastic partner in crime, Martin

13
00:00:51,040 --> 00:00:52,760
B. How are you Martin?

14
00:00:52,760 --> 00:00:54,560
I'm glad to be back.

15
00:00:54,560 --> 00:00:56,240
It's just nice to be back, isn't it?

16
00:00:56,240 --> 00:00:58,880
Yeah, everyone thought we'd forgot, but we hadn't.

17
00:00:58,880 --> 00:01:00,920
We're still here.

18
00:01:00,920 --> 00:01:01,920
Right.

19
00:01:01,920 --> 00:01:07,080
We are very sorry because we've been off air for probably as many as eight weeks now and

20
00:01:07,080 --> 00:01:08,600
a lot of cool stuff happened.

21
00:01:08,600 --> 00:01:11,840
And so make sure you sat down, get a coffee.

22
00:01:11,840 --> 00:01:15,600
We are going to go, oh, super speedy through a lot of different things that happened.

23
00:01:15,600 --> 00:01:18,240
We're not going to go through them in too much detail.

24
00:01:18,240 --> 00:01:22,420
They're probably old news now, such as the speed of things in AI, but some of the things

25
00:01:22,420 --> 00:01:23,640
are cool.

26
00:01:23,640 --> 00:01:24,640
What did we miss?

27
00:01:24,640 --> 00:01:31,120
Well, we missed figure one, the robotic company who partnered with open AI on a very interesting

28
00:01:31,120 --> 00:01:37,920
demo where they basically gave the figure robot open AI GP GPT-4 vision capabilities

29
00:01:37,920 --> 00:01:40,360
and text capabilities and speech capabilities.

30
00:01:40,360 --> 00:01:44,280
So basically the figure robot could see what was going on around it, interact with a human

31
00:01:44,280 --> 00:01:46,920
and basically respond to queries.

32
00:01:46,920 --> 00:01:47,920
It was kind of cool.

33
00:01:47,920 --> 00:01:48,920
That was pretty awesome.

34
00:01:48,920 --> 00:01:50,160
What else happened, mine?

35
00:01:50,160 --> 00:01:58,560
Well, we had Devon, which was a AI powered software engineer was previewed to much hype.

36
00:01:58,560 --> 00:02:07,360
It was able to autonomously plan and then execute a bunch of tasks, completing 14% of

37
00:02:07,360 --> 00:02:11,880
actions unassisted compared to about one or 2% of actions.

38
00:02:11,880 --> 00:02:17,680
However, when you scratch the surface slightly, a few people are somewhat skeptical with the

39
00:02:17,680 --> 00:02:22,440
performance and I don't think this is something that's going to be production ready in a hurry.

40
00:02:22,440 --> 00:02:28,920
But it was for a brief moment, at least an exciting look into the future of agentic.

41
00:02:28,920 --> 00:02:33,760
Agentic, mine got the word agentic straight in.

42
00:02:33,760 --> 00:02:37,280
He would just straight back in the mix and he's getting the word out.

43
00:02:37,280 --> 00:02:38,280
I love it.

44
00:02:38,280 --> 00:02:40,240
I know this is how we need to roll.

45
00:02:40,240 --> 00:02:43,320
Other things that happened is Claude three, right?

46
00:02:43,320 --> 00:02:44,820
Claude released a new model.

47
00:02:44,820 --> 00:02:47,760
They have a bunch of flavors, Haiku, Sonnet and Opus.

48
00:02:47,760 --> 00:02:53,560
Opus being the biggest model, super powerful and for a short while, at least took top model

49
00:02:53,560 --> 00:02:56,820
spot off of GPT-4 for many use cases.

50
00:02:56,820 --> 00:03:01,600
We played with it lots, Martin and I, we found it better at writing in terms of its stylistic

51
00:03:01,600 --> 00:03:04,840
capabilities and just to write more naturally like a human.

52
00:03:04,840 --> 00:03:09,360
It also was awesome at things like needle and haystack tests.

53
00:03:09,360 --> 00:03:12,000
It was really, really good at summarizing things like cool transcripts.

54
00:03:12,000 --> 00:03:15,360
Didn't really miss too many important details and was pretty awesome.

55
00:03:15,360 --> 00:03:19,760
But of course, there was some news this week, both from Google and from OpenAI that might

56
00:03:19,760 --> 00:03:20,880
surpass some of that.

57
00:03:20,880 --> 00:03:21,880
But we had Claude three.

58
00:03:21,880 --> 00:03:23,680
What else do we have, Martin?

59
00:03:23,680 --> 00:03:32,000
Hey, Jen, the AI avatar company that specializing lots of video production, they created a user

60
00:03:32,000 --> 00:03:39,240
generated content tool, meaning marketers can use Hey Jen avatar to create what looks

61
00:03:39,240 --> 00:03:41,200
to be user generated content.

62
00:03:41,200 --> 00:03:47,480
Now there is obviously a question around should you be using fake user generated content?

63
00:03:47,480 --> 00:03:50,280
That was an ethical question there.

64
00:03:50,280 --> 00:03:58,680
But I think if you want to overlay some customer reviews that are genuine and push them into

65
00:03:58,680 --> 00:04:05,580
video content that looks like real users working at home or in the office or move, speaking

66
00:04:05,580 --> 00:04:12,360
directly to camera, then Hey Jen with the UCG update can enable you to do that.

67
00:04:12,360 --> 00:04:16,520
That was actually really weird and cool because we've obviously shared quite a few Hey Jen

68
00:04:16,520 --> 00:04:22,520
things on, you know, when we're out on the speaking circuit or running conferences, running

69
00:04:22,520 --> 00:04:24,240
conferences, speaking conferences.

70
00:04:24,240 --> 00:04:28,620
And the videos are getting so much better now that even like the little bits that made

71
00:04:28,620 --> 00:04:32,400
it kind of a telltale that it was AI generated are starting to disappear, aren't they?

72
00:04:32,400 --> 00:04:36,320
They're starting to look much more like truly a human speaking.

73
00:04:36,320 --> 00:04:37,960
Yeah, very much so.

74
00:04:37,960 --> 00:04:44,920
Just the lip sync, all of those little elements that would give it away are greatly improved.

75
00:04:44,920 --> 00:04:46,920
Crumbs crazy.

76
00:04:46,920 --> 00:04:52,360
Other stuff we had Lama3 from Meta, they released the new version of their model.

77
00:04:52,360 --> 00:04:57,000
Interesting thing about this one is Lama3 also has three flavors, different size of

78
00:04:57,000 --> 00:05:00,680
models, in essence, different levels of capability.

79
00:05:00,680 --> 00:05:06,180
The smallest model is lightning fast, basically real time response.

80
00:05:06,180 --> 00:05:10,760
The middle model is the one that is the most powerful that they've released so far and

81
00:05:10,760 --> 00:05:17,160
is at least in some of our tests, almost as good as GPT-4 for certain aspects of things

82
00:05:17,160 --> 00:05:18,680
you might be trying to do.

83
00:05:18,680 --> 00:05:23,120
For example, summarizing call transcripts, but it's not even the biggest model, which

84
00:05:23,120 --> 00:05:24,300
they haven't released yet.

85
00:05:24,300 --> 00:05:27,320
So one has to wonder, they're still training it to be honest, it's one of the reasons we

86
00:05:27,320 --> 00:05:28,320
haven't seen it.

87
00:05:28,320 --> 00:05:33,200
But when it comes out, I think we can have expectations that it will be a GPT-4 killer

88
00:05:33,200 --> 00:05:39,240
if indeed OpenAI haven't killed their own GPT-4 with their own, you know, the king is

89
00:05:39,240 --> 00:05:41,160
dead long live the king from OpenAI.

90
00:05:41,160 --> 00:05:43,760
But that's something we'll talk about in a moment.

91
00:05:43,760 --> 00:05:45,960
Yeah, Lama3 is a really interesting one.

92
00:05:45,960 --> 00:05:53,080
It performs really well at lots of human interaction type tasks of answering questions and producing

93
00:05:53,080 --> 00:06:00,800
copy, but in the LLM chatbot arena, where it will often fall down is on things that

94
00:06:00,800 --> 00:06:08,320
require maths capabilities and isn't as good at say coding as the other models.

95
00:06:08,320 --> 00:06:10,120
What else do we have mine?

96
00:06:10,120 --> 00:06:16,260
We had Amazon Q. So Amazon has released a generative AI powered assistant that comes

97
00:06:16,260 --> 00:06:23,240
in two main versions, Amazon Q Business and Amazon Q Developer.

98
00:06:23,240 --> 00:06:27,600
Amazon Q Business is an AI powered chatbot that can answer questions, provide summaries,

99
00:06:27,600 --> 00:06:34,760
generate content based on data from an entire tech stack for enterprise systems.

100
00:06:34,760 --> 00:06:40,840
So taking data from Salesforce, Slack, Microsoft Office 365, things like that.

101
00:06:40,840 --> 00:06:46,240
And Amazon Q Developer is something that obviously assists software developers and IT.

102
00:06:46,240 --> 00:06:51,680
The professional so it can help with coding, debugging, testing and has some multi step

103
00:06:51,680 --> 00:06:58,580
reasoning capabilities, plugs into 40 different enterprise data sources and has all of the

104
00:06:58,580 --> 00:07:03,480
enterprise grade security functions that you would expect.

105
00:07:03,480 --> 00:07:04,480
Yes.

106
00:07:04,480 --> 00:07:09,040
We also have Microsoft news that Microsoft is developing its own new large model called

107
00:07:09,040 --> 00:07:15,200
MAI1, which is going to be led by Mustafa Suleiman, who was the CEO of Inflection, which

108
00:07:15,200 --> 00:07:20,200
created the product PI that we really like for a while until Microsoft basically gutted

109
00:07:20,200 --> 00:07:27,800
PI by bringing all of their stuff over to Microsoft and effectively almost purchasing

110
00:07:27,800 --> 00:07:33,960
Inflection and the company behind it by buying all of their graphics cards that they got

111
00:07:33,960 --> 00:07:36,440
from Nvidia to train models on.

112
00:07:36,440 --> 00:07:38,600
That's probably an oversimplified summary.

113
00:07:38,600 --> 00:07:42,480
But basically, Microsoft went, let's go get some more talent and let's build our own

114
00:07:42,480 --> 00:07:43,480
model.

115
00:07:43,480 --> 00:07:48,840
And that's interesting in the context of Microsoft's partnership with OpenAI at the same time as

116
00:07:48,840 --> 00:07:54,280
OpenAI appears to be dancing rather close at the party with Apple, maybe even whispering

117
00:07:54,280 --> 00:07:59,500
some sweet nothings into Apple's ear around putting their latest models and their technology

118
00:07:59,500 --> 00:08:00,560
into Siri.

119
00:08:00,560 --> 00:08:05,840
So there's a bunch of uneasy partnerships forming and it's interesting to see that Microsoft

120
00:08:05,840 --> 00:08:09,600
is now working on its own large model.

121
00:08:09,600 --> 00:08:14,040
Watch this space in terms of figuring out how that is going to play out.

122
00:08:14,040 --> 00:08:23,400
As well as that, we had the first commissioned music video produced using OpenAI's Sora.

123
00:08:23,400 --> 00:08:29,320
So I actually don't know the name of the band that did this, but it was quite an interesting

124
00:08:29,320 --> 00:08:30,760
production process.

125
00:08:30,760 --> 00:08:35,280
And if you've been watching, I don't know, Paul, if you've read any reports of people

126
00:08:35,280 --> 00:08:41,800
that have been using Sora and trying to use it as a creative tool, it sounds like it's

127
00:08:41,800 --> 00:08:47,800
as easy and straightforward to use as using the early versions of Mid Journey.

128
00:08:47,800 --> 00:08:56,320
Yeah, I think it was an independent artist called Washed Out and the single was called

129
00:08:56,320 --> 00:08:58,080
The Hardest Part.

130
00:08:58,080 --> 00:09:00,160
It is interesting.

131
00:09:00,160 --> 00:09:07,160
Today I think I sent you an article on the WhatsApps about one of the preliminary testers

132
00:09:07,160 --> 00:09:15,720
of OpenAI's Sora video model coming clean that their impressive video with the balloon

133
00:09:15,720 --> 00:09:22,760
head was not quite as AI generated as it appears because the way that that example was positioned.

134
00:09:22,760 --> 00:09:26,560
And I think we even talked about this a couple of months ago in our last series of episodes.

135
00:09:26,560 --> 00:09:32,800
I think we had an episode on Sora if I remember rightly, but it looked like the whole thing

136
00:09:32,800 --> 00:09:38,120
was obviously edited by a human, but had been generated by Sora.

137
00:09:38,120 --> 00:09:44,320
But actually, no, sometimes it generated a floating balloon above a human and they had

138
00:09:44,320 --> 00:09:50,160
to like rotoscope out the human's head, which is not how it was made to appear.

139
00:09:50,160 --> 00:09:53,560
So yes, Sora is early Mid Journey.

140
00:09:53,560 --> 00:09:54,560
It does cool stuff.

141
00:09:54,560 --> 00:09:58,560
It almost sounds extremely difficult to control and get what you want out of it.

142
00:09:58,560 --> 00:10:01,520
Obviously most of us don't even have the chance to play with it yet.

143
00:10:01,520 --> 00:10:02,920
So I think you're right.

144
00:10:02,920 --> 00:10:07,520
I think it's video is a really cool area that's improving very quickly, but our ability to

145
00:10:07,520 --> 00:10:11,960
actually get interesting stuff that we want out of it is probably a couple of years away

146
00:10:11,960 --> 00:10:14,780
if Mid Journey is anything to go by.

147
00:10:14,780 --> 00:10:16,660
And then what else do we have?

148
00:10:16,660 --> 00:10:19,440
We had Alpha Fold 3.

149
00:10:19,440 --> 00:10:24,880
So I'm obviously a biologically nerd working in life science marketing and Google Deep

150
00:10:24,880 --> 00:10:28,240
Mind bringing out their Alpha Fold 3 is pretty cool.

151
00:10:28,240 --> 00:10:34,400
So Alpha Fold 3 is an improvement on Alpha Fold 2 because now I can predict the structures

152
00:10:34,400 --> 00:10:40,580
of not just proteins, but also DNA, RNA, ligands, and also when they are interacting.

153
00:10:40,580 --> 00:10:48,200
So the goal is to move ever closer to being able to simulate biological systems in a computer

154
00:10:48,200 --> 00:10:53,320
so that you can predict how different things will interact to help you identify interesting

155
00:10:53,320 --> 00:11:00,000
new drug targets, but also more rapidly and easily digitally screen for drug candidates

156
00:11:00,000 --> 00:11:03,160
that might interact with those things and get the outcome that you want, whether that's

157
00:11:03,160 --> 00:11:08,200
trying to treat Alzheimer's or reduce inflammation or treat cancer.

158
00:11:08,200 --> 00:11:11,880
Is a lot of excitement about this and I'm very excited about it as well, but there is

159
00:11:11,880 --> 00:11:16,080
still a need to validate what the computer thinks through wet lab research and then put

160
00:11:16,080 --> 00:11:17,400
those things in clinical trials.

161
00:11:17,400 --> 00:11:23,440
So I think if you hear too much hyperbole around how this is going to impact the drug

162
00:11:23,440 --> 00:11:26,200
discovery process, it is an awesome tool.

163
00:11:26,200 --> 00:11:30,600
I think it will speed up basic research a huge amount and I think it will also help

164
00:11:30,600 --> 00:11:36,240
with huge amounts of the drug discovery process and chop some of the time it takes to find

165
00:11:36,240 --> 00:11:37,240
interesting drugs.

166
00:11:37,240 --> 00:11:41,440
But there's still some stuff that you have to do in the real world, unfortunately, as

167
00:11:41,440 --> 00:11:47,600
it stands, but it's still a big leap forward and it will be super cool to see how people

168
00:11:47,600 --> 00:11:48,880
play with it.

169
00:11:48,880 --> 00:11:52,000
Thank you for letting me nerd out a little bit on that one for a while, Martin.

170
00:11:52,000 --> 00:11:54,600
Well, these things are only going to get better from here on it, right?

171
00:11:54,600 --> 00:11:58,040
So Alpha Fold 2 was what last year?

172
00:11:58,040 --> 00:12:01,560
Maybe even older than that, I would say, maybe 2022.

173
00:12:01,560 --> 00:12:02,560
2022.

174
00:12:02,560 --> 00:12:05,720
Well, yeah, I think so.

175
00:12:05,720 --> 00:12:06,960
It's pretty recent, right?

176
00:12:06,960 --> 00:12:15,000
And this is a big step change in a very exciting domain.

177
00:12:15,000 --> 00:12:16,000
Crumbs.

178
00:12:16,000 --> 00:12:17,560
Well, we may even have digital twins.

179
00:12:17,560 --> 00:12:19,920
We might be simulating humans at that point.

180
00:12:19,920 --> 00:12:23,120
And then, of course, being able to figure out how drugs are going to interact with things

181
00:12:23,120 --> 00:12:26,360
when you can simulate maybe a brain or a heart now.

182
00:12:26,360 --> 00:12:30,600
We're quite into interesting territory in terms of what you might need to do in the

183
00:12:30,600 --> 00:12:33,480
real world versus what you can do in the old computer.

184
00:12:33,480 --> 00:12:36,640
Well, let's do we see Martin one more.

185
00:12:36,640 --> 00:12:40,560
We saw the takedown.

186
00:12:40,560 --> 00:12:48,560
It's the only way of some well-hyped AI hardware.

187
00:12:48,560 --> 00:12:55,760
So the Humane Pin and the Rabbit R1, which were two pieces of AI powered hardware that

188
00:12:55,760 --> 00:13:04,000
had been previewed many months ago, the first editions of them got into the hands of tech

189
00:13:04,000 --> 00:13:11,880
reviewers, MKBHD, one of the biggest reviewers of tech on YouTube, does fantastic reviews.

190
00:13:11,880 --> 00:13:16,040
He was one of the first to put out some reviews of both of these tools.

191
00:13:16,040 --> 00:13:20,080
And yet they didn't didn't land well.

192
00:13:20,080 --> 00:13:21,280
They're half baked.

193
00:13:21,280 --> 00:13:24,600
They don't really do anything that you can't do on a smartphone.

194
00:13:24,600 --> 00:13:28,080
The user interface for them is just clunky.

195
00:13:28,080 --> 00:13:33,800
Everything about them is just hard to use and not at all appealing for anyone.

196
00:13:33,800 --> 00:13:35,760
So yeah, what's this space?

197
00:13:35,760 --> 00:13:43,560
I haven't had any interest in buying either of these, but I have pre-ordered the Limitless

198
00:13:43,560 --> 00:13:49,760
Pendant, which is a piece of AI hardware, and that should be arriving within a matter

199
00:13:49,760 --> 00:13:50,760
of months.

200
00:13:50,760 --> 00:13:55,440
So hopefully that is slightly better.

201
00:13:55,440 --> 00:14:00,400
Given the reports that came in about these pieces of tech, I wouldn't hold your breath.

202
00:14:00,400 --> 00:14:06,320
Although obviously what you have acquired is cheaper and simpler and much more likely

203
00:14:06,320 --> 00:14:09,240
to work, I would think.

204
00:14:09,240 --> 00:14:15,640
We haven't had time to talk about the upgrades to Meta's Ray-Ban glasses that have a mic

205
00:14:15,640 --> 00:14:18,120
and a camera that are becoming AI enabled.

206
00:14:18,120 --> 00:14:20,700
So in essence, you can talk to them.

207
00:14:20,700 --> 00:14:26,680
And if you think about some of the sort of 10 net, I can say it in a minute, technological

208
00:14:26,680 --> 00:14:32,100
leaps that we've seen this week, that form factor speaking to your glasses while they

209
00:14:32,100 --> 00:14:36,920
can see the world around you, even before we get to technology that can overlay information

210
00:14:36,920 --> 00:14:42,680
in your view, which of course is not currently possible in that type of simple glasses form

211
00:14:42,680 --> 00:14:45,520
factor is actually going to be quite interesting.

212
00:14:45,520 --> 00:14:50,160
So while some of the early technology hasn't quite panned out, I think we're still just

213
00:14:50,160 --> 00:14:53,680
trying to figure out what's the right form factor for different applications.

214
00:14:53,680 --> 00:14:58,800
But you said are we better off having apps on our phones, apps on our watches, smarter

215
00:14:58,800 --> 00:15:00,320
glasses?

216
00:15:00,320 --> 00:15:06,840
What I think Rabbit and the Humane Pin have shown is probably those form factors are not

217
00:15:06,840 --> 00:15:07,840
that great.

218
00:15:07,840 --> 00:15:09,880
Although I think people have had fun playing with the Rabbit.

219
00:15:09,880 --> 00:15:15,720
It just couldn't do 90% of the stuff that it was touted to be able to do.

220
00:15:15,720 --> 00:15:17,280
Maybe it will be able to at some point.

221
00:15:17,280 --> 00:15:21,640
And they're both really slow at responding, which of course is a very frustrating user

222
00:15:21,640 --> 00:15:22,640
experience.

223
00:15:22,640 --> 00:15:29,680
So maybe we can use response times nine as a little cheeky segue into one of the bigger

224
00:15:29,680 --> 00:15:33,840
weeks that we've had in AI this year, because we're going to spend the most of the rest

225
00:15:33,840 --> 00:15:39,640
of this episode talking about open AI releasing GPT-4-0.

226
00:15:39,640 --> 00:15:45,800
Yes, these model names are a pain to try and remember and pronounce.

227
00:15:45,800 --> 00:15:50,640
Please hire a branding expert at some point, someone.

228
00:15:50,640 --> 00:15:55,800
And also Google I.O. conference this week, developer conference, where they also announced

229
00:15:55,800 --> 00:16:00,640
a number of significant upgrades to some of the technologies they've released recently

230
00:16:00,640 --> 00:16:05,820
and also gave us a sneak peek of where their technologies, their AI enabled technologies

231
00:16:05,820 --> 00:16:11,600
for business and of course marketing, where they expect them to go in the future.

232
00:16:11,600 --> 00:16:16,480
To kick things off, Martin, why don't you tell us what open AI have been up to this

233
00:16:16,480 --> 00:16:17,480
week?

234
00:16:17,480 --> 00:16:23,320
Well, they had a live streamed event from their offices where they announced, as you

235
00:16:23,320 --> 00:16:27,000
say, GPT-4-0.

236
00:16:27,000 --> 00:16:29,720
O stands for Omni.

237
00:16:29,720 --> 00:16:35,180
This is their multi-modal GPT-4 version.

238
00:16:35,180 --> 00:16:39,240
So it's fully multi-modal within the same neural network.

239
00:16:39,240 --> 00:16:47,440
So what that means is text input, text output, image input, image output, which is what we

240
00:16:47,440 --> 00:16:52,880
really interesting and we'll talk more about that in a moment, audio input and audio output.

241
00:16:52,880 --> 00:17:00,800
So if you've been using the chat GPT app on mobile, you've probably tried the voice input

242
00:17:00,800 --> 00:17:04,760
where you can try speaking to it and it will talk back to you.

243
00:17:04,760 --> 00:17:10,840
That was a clever little piece of software engineering by OpenAI where what they did

244
00:17:10,840 --> 00:17:14,400
was they plugged multiple models together.

245
00:17:14,400 --> 00:17:21,320
So you had OpenAI whisper doing the transcription for your voice and then they had their own

246
00:17:21,320 --> 00:17:30,680
text to voice service transcribe open eyes, GPT-4's responses and speak it back to you.

247
00:17:30,680 --> 00:17:34,960
So it was basically cobbling together multiple models to give you the impression of this

248
00:17:34,960 --> 00:17:37,380
fully interactive experience.

249
00:17:37,380 --> 00:17:44,440
Now using GPT-4 O, all of that is within the one model.

250
00:17:44,440 --> 00:17:50,400
So the experience for the user will be much quicker.

251
00:17:50,400 --> 00:17:59,240
The model itself, GPT-4 is as good in terms of quality outputs as GPT-4, so they claim,

252
00:17:59,240 --> 00:18:04,840
but they're able to offer it at a 50% price reduction, meaning they've made it incredibly

253
00:18:04,840 --> 00:18:14,080
efficient, much more compute light, so it doesn't use as much resource per inference.

254
00:18:14,080 --> 00:18:19,820
So every time you use it, it just isn't using as much energy and computing power.

255
00:18:19,820 --> 00:18:25,720
That's one of the biggest developments really is that they've managed to create this model,

256
00:18:25,720 --> 00:18:28,840
which is so much more efficient.

257
00:18:28,840 --> 00:18:41,400
And what that has enabled them to do is release GPT-4 O for free to all chat GPT users.

258
00:18:41,400 --> 00:18:49,520
So where free chat GPT users previously only had access to GPT 3.5, which as we know is

259
00:18:49,520 --> 00:18:53,240
at this stage, not really a great model.

260
00:18:53,240 --> 00:18:54,240
It's pants.

261
00:18:54,240 --> 00:18:56,040
It's pants, Mike.

262
00:18:56,040 --> 00:18:59,000
Yes, it's pants.

263
00:18:59,000 --> 00:19:06,800
Now everybody gets this state of the art model for free.

264
00:19:06,800 --> 00:19:11,760
And they also get that multimodality experience as well.

265
00:19:11,760 --> 00:19:14,480
This isn't actually launched for everybody just yet.

266
00:19:14,480 --> 00:19:15,480
This is coming.

267
00:19:15,480 --> 00:19:20,360
It's a few weeks away from being fully rolled out and deployed worldwide.

268
00:19:20,360 --> 00:19:22,080
But this is what users can expect to see.

269
00:19:22,080 --> 00:19:28,280
And you've got to think that this is going to be huge for open AI because there are reports

270
00:19:28,280 --> 00:19:39,520
that I've seen that suggest chat GPT stickiness isn't very high, that people will sign up,

271
00:19:39,520 --> 00:19:45,000
use it once and then go, oh, yeah, that was kind of interesting.

272
00:19:45,000 --> 00:19:47,520
But it was a little bit disappointing.

273
00:19:47,520 --> 00:19:50,160
And that's probably because they were using 3.5.

274
00:19:50,160 --> 00:19:56,520
But now everybody's going to be able to get the state of the art intelligence for free.

275
00:19:56,520 --> 00:20:00,640
I think this is critical, Martin, because how many people do we speak with who don't

276
00:20:00,640 --> 00:20:07,760
have a pro account, who believe that the quality of large language models is what GPT 3.5 can

277
00:20:07,760 --> 00:20:08,760
do?

278
00:20:08,760 --> 00:20:10,520
They use chat GPT because it's the one they've heard of.

279
00:20:10,520 --> 00:20:12,960
They haven't even thought to go and have a play with Claude because they just haven't

280
00:20:12,960 --> 00:20:14,080
heard of it.

281
00:20:14,080 --> 00:20:17,800
And are left thinking, oh, AI, why is everybody excited about AI?

282
00:20:17,800 --> 00:20:22,160
It sucks when actually it's GPT 3.5 that sucks.

283
00:20:22,160 --> 00:20:24,480
And 4.0 doesn't suck.

284
00:20:24,480 --> 00:20:25,680
In fact, it's rather awesome.

285
00:20:25,680 --> 00:20:28,280
So I think this will be the opportunity for people.

286
00:20:28,280 --> 00:20:34,480
I hope people go back and play again and are like, oh, OK, this is why people were impressed.

287
00:20:34,480 --> 00:20:36,800
And there are other models coming soon.

288
00:20:36,800 --> 00:20:38,680
We know this is true.

289
00:20:38,680 --> 00:20:43,560
But I do think this will be enough to make people go, actually, I can perform some work

290
00:20:43,560 --> 00:20:45,080
functions with this.

291
00:20:45,080 --> 00:20:47,600
So I think that power in itself is pretty exciting.

292
00:20:47,600 --> 00:20:54,080
Before we even get into the expanded capabilities of what multimodal, omnimodal, whatever they're

293
00:20:54,080 --> 00:20:55,960
calling it, what it enables.

294
00:20:55,960 --> 00:20:56,960
Yeah.

295
00:20:56,960 --> 00:20:59,640
So let's go straight into those, shall we?

296
00:20:59,640 --> 00:21:08,060
The first big thing that they really previewed was the voice capabilities.

297
00:21:08,060 --> 00:21:15,480
And if you ever used chat GPT voice mode previously, you'll know that it was a little bit laggy.

298
00:21:15,480 --> 00:21:22,280
There was a three or four second delay between you speaking into the model and then getting

299
00:21:22,280 --> 00:21:24,920
a response back.

300
00:21:24,920 --> 00:21:26,600
That's because there were multiple steps in it.

301
00:21:26,600 --> 00:21:31,480
As I explained earlier, it was describing, it was then sending it off and doing a bunch

302
00:21:31,480 --> 00:21:34,880
of other processes with multiple models.

303
00:21:34,880 --> 00:21:40,880
Now because it's all within the same model, it's much, much quicker.

304
00:21:40,880 --> 00:21:46,560
So having a conversation in voice mode, which is just like having a conversation on the

305
00:21:46,560 --> 00:21:55,920
phone, the model speaks back to you with a typical delay of around point two to point

306
00:21:55,920 --> 00:21:58,820
three seconds.

307
00:21:58,820 --> 00:22:02,960
So a third of a second delay quicker than a human.

308
00:22:02,960 --> 00:22:05,120
It feels like that.

309
00:22:05,120 --> 00:22:06,920
It's so quick.

310
00:22:06,920 --> 00:22:11,360
But also the voice, the voice Paul.

311
00:22:11,360 --> 00:22:12,520
Crumbs, the voice.

312
00:22:12,520 --> 00:22:17,840
I mean, Sam Altman and a bunch of other people were just basically tweeting the word her,

313
00:22:17,840 --> 00:22:23,760
which for those that have seen the movie, the 2013, I think movie about AI companions,

314
00:22:23,760 --> 00:22:30,200
AI assistants that speak and sound and empathize like humans, but basically live in your ear

315
00:22:30,200 --> 00:22:35,080
as an earbud speaking to you all day and they can see what you see because of a ironically

316
00:22:35,080 --> 00:22:41,520
humane AI like pin or perhaps the equivalent, I guess, of putting your phone in your top

317
00:22:41,520 --> 00:22:46,400
pocket of your shirt or your jacket.

318
00:22:46,400 --> 00:22:52,320
That's kind of been sci-fi pinnacle of, oh, wouldn't it be cool if you had the opportunity

319
00:22:52,320 --> 00:23:02,040
to interact with a computer like a human, but crumbs, this is pretty close to that.

320
00:23:02,040 --> 00:23:05,640
The female voice sounds like it's like it laughs.

321
00:23:05,640 --> 00:23:09,640
It kind of almost is borderline flirty, which I think some people have found like a bit

322
00:23:09,640 --> 00:23:10,640
creepy.

323
00:23:10,640 --> 00:23:15,880
But if you think like humans show a lot of dynamism in how they communicate, right, they

324
00:23:15,880 --> 00:23:22,120
do laugh, they do tease, they do emote, I guess is the word.

325
00:23:22,120 --> 00:23:25,160
And this feels like a massive jump forward to me.

326
00:23:25,160 --> 00:23:29,040
Like I can't believe that more people are not speaking about it because it's like a

327
00:23:29,040 --> 00:23:30,040
human.

328
00:23:30,040 --> 00:23:41,000
Yeah, it's because the model itself has the ability to understand voice and it isn't using

329
00:23:41,000 --> 00:23:44,160
a text to speech to generate the voice.

330
00:23:44,160 --> 00:23:48,040
Like you say, it can emote.

331
00:23:48,040 --> 00:23:51,400
It can, if you ask it to speak faster, it will speak faster.

332
00:23:51,400 --> 00:23:53,360
If you ask it to slow down, it will slow down.

333
00:23:53,360 --> 00:23:55,520
If you ask it to whisper, it will whisper.

334
00:23:55,520 --> 00:23:58,120
It can do all of that within the voice.

335
00:23:58,120 --> 00:24:06,800
And also, in the input of the voice, when you speak to it, it can detect those emotional

336
00:24:06,800 --> 00:24:07,800
ranges.

337
00:24:07,800 --> 00:24:10,960
So if you are excited, it will detect that you're excited.

338
00:24:10,960 --> 00:24:14,320
If you're sad, it will pick up on the fact that you're sad.

339
00:24:14,320 --> 00:24:19,200
If you sound somewhat confused or sound a little bit sarcastic, it will pick up on all

340
00:24:19,200 --> 00:24:20,240
of these things.

341
00:24:20,240 --> 00:24:21,240
It's crazy.

342
00:24:21,240 --> 00:24:25,480
And people, I don't think, same as you say there, I think people are sleeping on this

343
00:24:25,480 --> 00:24:34,160
slightly, because this capability is, yeah, I mean, it takes AI to another level.

344
00:24:34,160 --> 00:24:38,640
And we should also say that in a number of the demos, you have the demos during the live

345
00:24:38,640 --> 00:24:43,400
event, but there's also a bunch of videos on the OpenAI site, which I recommend everybody

346
00:24:43,400 --> 00:24:46,080
interested in this goes and watches.

347
00:24:46,080 --> 00:24:51,020
So we've got humans interacting with chat GPT in essence, speaking to it, it speaks

348
00:24:51,020 --> 00:24:52,020
back in near real time.

349
00:24:52,020 --> 00:24:54,040
So it's like having a conversation with a human.

350
00:24:54,040 --> 00:24:59,800
But you can also screen grab images or take pictures on your phone and send those and

351
00:24:59,800 --> 00:25:07,120
effectively live stream video to the app so it can see what you see or what you're doing.

352
00:25:07,120 --> 00:25:09,800
And this just opens up a huge amount of use cases.

353
00:25:09,800 --> 00:25:15,360
So one of the example videos, somebody gets two phones and they let one version of chat

354
00:25:15,360 --> 00:25:17,680
GPT speak to another version of chat GPT.

355
00:25:17,680 --> 00:25:22,320
And it mimics the customer service conversation about returning an iPhone.

356
00:25:22,320 --> 00:25:28,680
And the whole thing happens without any human input right up to the end when one of the

357
00:25:28,680 --> 00:25:33,720
chat GPT asks for the person's email address so that they can send over details on how

358
00:25:33,720 --> 00:25:35,440
to do the return shipping.

359
00:25:35,440 --> 00:25:42,040
And the other chat GPT says, please hold and then goes, Bob, check your email.

360
00:25:42,040 --> 00:25:45,520
Did you get an email from the service provider or something similar to that?

361
00:25:45,520 --> 00:25:48,320
And it was like, wow, this is not far away.

362
00:25:48,320 --> 00:25:53,560
The customer service on the phone is very, very close to getting the level of disruption

363
00:25:53,560 --> 00:25:55,780
that we all predicted was going to come.

364
00:25:55,780 --> 00:26:01,560
In another example, we had the team from the Khan Academy showing a demonstration of how

365
00:26:01,560 --> 00:26:04,920
the tool could be used as a personal coach and tutor.

366
00:26:04,920 --> 00:26:09,360
And in this example, using the new desktop app, which has also been released, and it

367
00:26:09,360 --> 00:26:15,320
looks like it will be on iPad and Apple devices to begin with, not Windows.

368
00:26:15,320 --> 00:26:19,440
We'll be looking back at some of the friction and tension that we talked about at the beginning.

369
00:26:19,440 --> 00:26:26,000
But in this particular example, a teenager is being taken through a fairly complex math

370
00:26:26,000 --> 00:26:31,640
problem, you know, teenage 14 year old mathematics problem around right-angled triangles and

371
00:26:31,640 --> 00:26:33,100
what have you.

372
00:26:33,100 --> 00:26:39,040
And because chat GPT can both have a conversation with the person, but also see what's on their

373
00:26:39,040 --> 00:26:44,560
screen, it can ask questions and provide feedback to the person about what they're trying to

374
00:26:44,560 --> 00:26:46,640
do.

375
00:26:46,640 --> 00:26:53,440
And this is fundamentally mind blowing because we Martin and I were talking off air.

376
00:26:53,440 --> 00:26:56,360
This fundamentally changes how these tools can interact with us.

377
00:26:56,360 --> 00:27:01,960
If chat GPT can see your desktop while you're working, you can ask it questions when you

378
00:27:01,960 --> 00:27:06,680
get stuck and it can be constantly watching what you do, obviously privacy and a number

379
00:27:06,680 --> 00:27:09,120
of data issues around how we feel about that.

380
00:27:09,120 --> 00:27:13,760
But basically coaching you on how to do things well, like I'm written an email, I'm about

381
00:27:13,760 --> 00:27:18,160
to send it before I do chat GPT just says, Paul, just so you know, that last statement

382
00:27:18,160 --> 00:27:19,160
is a bit confusing.

383
00:27:19,160 --> 00:27:22,280
You might want to consider rewording it because other people might not be able to understand

384
00:27:22,280 --> 00:27:23,280
it.

385
00:27:23,280 --> 00:27:24,720
Do you want me to have a go at rewording it for you?

386
00:27:24,720 --> 00:27:25,720
Yes, please.

387
00:27:25,720 --> 00:27:30,600
Chat GPT or maybe I don't know if you realize Paul, but that last statement could be seen

388
00:27:30,600 --> 00:27:31,600
as a bit inflammatory.

389
00:27:31,600 --> 00:27:33,600
Do you want to soften it down for this email?

390
00:27:33,600 --> 00:27:35,080
Yes, chat GPT.

391
00:27:35,080 --> 00:27:38,200
I think that would be a good idea.

392
00:27:38,200 --> 00:27:42,160
Not least part of the announcement that seems to have gone live today is the ability of

393
00:27:42,160 --> 00:27:46,320
chat GPT to natively integrate with Google Drive and Microsoft OneDrive files.

394
00:27:46,320 --> 00:27:50,200
So we can actually pull information from your existing system.

395
00:27:50,200 --> 00:27:53,080
We're probably not that far away from it doing that dynamically at the moment.

396
00:27:53,080 --> 00:27:55,820
You have to tell it what file you want to work with.

397
00:27:55,820 --> 00:27:59,220
So it can say things like, Paul, do you think it would be useful to attach this thing to

398
00:27:59,220 --> 00:28:02,040
this email, given that it's about this topic?

399
00:28:02,040 --> 00:28:04,840
Like this is how the world is going to be very soon.

400
00:28:04,840 --> 00:28:07,320
I think Martin and it's.

401
00:28:07,320 --> 00:28:12,880
I think we've got the potential for all of us to have at work and AI coach helping us

402
00:28:12,880 --> 00:28:19,040
get better and better at what we do with personalized coaching, which I think would be amazing.

403
00:28:19,040 --> 00:28:24,400
But the downside of that is once it gets good enough to coach us on a lot of these activities,

404
00:28:24,400 --> 00:28:28,840
all it needs is the power to be able to move your mouse and click for you.

405
00:28:28,840 --> 00:28:31,960
And then it's do you need the human to be doing the things at all?

406
00:28:31,960 --> 00:28:33,880
And I guess that's where it gets a bit draconian.

407
00:28:33,880 --> 00:28:38,080
And we were talking offline, this is kind of weird.

408
00:28:38,080 --> 00:28:39,080
Like where's this going to head out?

409
00:28:39,080 --> 00:28:40,080
Head out.

410
00:28:40,080 --> 00:28:41,520
I don't know what you think about some of the stuff, Mark.

411
00:28:41,520 --> 00:28:42,520
It's pretty mind blowing.

412
00:28:42,520 --> 00:28:44,800
I mean, it still makes mistakes.

413
00:28:44,800 --> 00:28:52,560
So I think that's that's going to work in the human's favor for a while until it is

414
00:28:52,560 --> 00:28:59,480
reliable to such a degree that you can you can just chuck it a task and know it's going

415
00:28:59,480 --> 00:29:00,920
to get done at the moment.

416
00:29:00,920 --> 00:29:01,920
It does still make mistakes.

417
00:29:01,920 --> 00:29:05,240
I've tried it with some coding tasks.

418
00:29:05,240 --> 00:29:09,480
And I mean, before this call, I was saying it's amazing.

419
00:29:09,480 --> 00:29:18,920
You can give it a solid brief and it will perform full one shot coding outputs, complete

420
00:29:18,920 --> 00:29:22,640
and whole that work brilliantly better than GPT-4.

421
00:29:22,640 --> 00:29:26,040
So it's definitely I feel it's a step on from that.

422
00:29:26,040 --> 00:29:31,320
And I've seen other people in the community that pushing these models saying the same

423
00:29:31,320 --> 00:29:33,360
thing as well.

424
00:29:33,360 --> 00:29:35,560
But then it still does make some mistakes.

425
00:29:35,560 --> 00:29:40,760
And I've had it spit out a few errors in various instances.

426
00:29:40,760 --> 00:29:44,280
So I think we're safe from that for a while yet.

427
00:29:44,280 --> 00:29:47,360
It's funny you mentioned about having it as an assistant.

428
00:29:47,360 --> 00:29:52,120
Sam Altman did an interview on I can't remember which podcast it was now, but straight after

429
00:29:52,120 --> 00:29:58,880
the announcement, this podcast was published on this other feed.

430
00:29:58,880 --> 00:30:04,320
And Sam Altman was asked how he's been using this new model.

431
00:30:04,320 --> 00:30:08,680
He said he's only had access to it for a week.

432
00:30:08,680 --> 00:30:10,080
Whether that's true or not.

433
00:30:10,080 --> 00:30:14,000
But he he described his interactions with it.

434
00:30:14,000 --> 00:30:17,040
And he said what he does is he has it the conversation.

435
00:30:17,040 --> 00:30:22,760
It just has the chat mode open on his desk while he's doing some work.

436
00:30:22,760 --> 00:30:26,760
And it's just there sitting there, always listening.

437
00:30:26,760 --> 00:30:33,080
And rather than flicking between tabs to find out information or to go and do things that

438
00:30:33,080 --> 00:30:36,760
he would normally kind of second screen.

439
00:30:36,760 --> 00:30:39,000
What he's doing is he's just asking this.

440
00:30:39,000 --> 00:30:43,280
He's just speaking to it as he's going and he's finding information as he goes and it's

441
00:30:43,280 --> 00:30:46,120
pulling back responses to him.

442
00:30:46,120 --> 00:30:48,960
And it's helping him not have to second screen.

443
00:30:48,960 --> 00:30:52,280
It's just like a like an assistant, right?

444
00:30:52,280 --> 00:30:54,160
That's what it's like.

445
00:30:54,160 --> 00:30:55,560
That's how it's been positioned.

446
00:30:55,560 --> 00:30:59,320
Yeah, and he I think there was another podcast and I'm going to paraphrase here.

447
00:30:59,320 --> 00:31:02,920
So hopefully I don't get I don't drift too far away from what he originally said.

448
00:31:02,920 --> 00:31:08,960
But he was he was being asked like what does he think these systems like what does he want

449
00:31:08,960 --> 00:31:10,800
these systems to be able to do?

450
00:31:10,800 --> 00:31:15,520
I think it was on the All In podcast actually on Saturday of Friday it was released.

451
00:31:15,520 --> 00:31:24,720
And in essence what he wants is in a business context is a highly capable executive assistant

452
00:31:24,720 --> 00:31:29,280
that can do things on his behalf but that can really act independently because it has

453
00:31:29,280 --> 00:31:34,800
an ability to understand the context of the problems that he's dealing with and the tasks

454
00:31:34,800 --> 00:31:41,760
he's trying to execute and can support them thoughtfully and wherever possible autonomously.

455
00:31:41,760 --> 00:31:44,360
So it's clear that that's the vision of where they're trying to head.

456
00:31:44,360 --> 00:31:49,000
They keep telling us that there's been a couple of podcasts recently that GPT-4 kind of sucks

457
00:31:49,000 --> 00:31:54,000
and yeah GPT-4 is an improvement and it has some really cool stuff but it suggests that

458
00:31:54,000 --> 00:31:57,240
some of the other things they're working on might imply that even some of the aspects

459
00:31:57,240 --> 00:31:59,640
of these models kind of suck as well.

460
00:31:59,640 --> 00:32:01,900
So I think you're right.

461
00:32:01,900 --> 00:32:07,800
I do think the capacity for mistakes is still fairly large and a human in the loop is absolutely

462
00:32:07,800 --> 00:32:08,940
critical.

463
00:32:08,940 --> 00:32:15,200
I would love to see us discussing GPT-5 which is rumored before the end of the summer but

464
00:32:15,200 --> 00:32:20,460
we shall see how much of an improvement it brings how much improvement it brings in terms

465
00:32:20,460 --> 00:32:23,000
of its agenticity.

466
00:32:23,000 --> 00:32:26,600
A little riff on your agentic there mate.

467
00:32:26,600 --> 00:32:33,560
Because of course if we can reliably perform multi-step tasks, reason within those tasks

468
00:32:33,560 --> 00:32:38,520
and now we have this ability to show it what we're doing, speak to it in real time, that's

469
00:32:38,520 --> 00:32:43,000
going to be really really interesting and to finish the thought on the coaches versus

470
00:32:43,000 --> 00:32:48,000
the AI does it for us, it'll be really interesting to see how it plays out because one of the

471
00:32:48,000 --> 00:32:52,580
things you and I have discussed is will there be policies that are put in place that says

472
00:32:52,580 --> 00:32:58,340
AI can be a coach for certain roles and it can coach a human to do them but it can't

473
00:32:58,340 --> 00:33:00,480
do the work itself.

474
00:33:00,480 --> 00:33:06,280
To maintain, A, because that might be safer in a lot of roles where we just can't afford

475
00:33:06,280 --> 00:33:16,060
for the AI to make a mistake but also to ensure that some level of human first AI is maintained

476
00:33:16,060 --> 00:33:18,660
because that might be needed as it gets better and better and better.

477
00:33:18,660 --> 00:33:25,920
So it's super fascinating, quite impressive and I'm personally looking forward to it

478
00:33:25,920 --> 00:33:31,640
being released because if you go into Chatchipity now and you've got a paid account you'll

479
00:33:31,640 --> 00:33:36,180
see you can access 4.0 already so you can take advantage of a number of the multimodal

480
00:33:36,180 --> 00:33:39,840
things that Martin described but what you won't be able to do yet is speak with it because

481
00:33:39,840 --> 00:33:42,040
that hasn't been released yet.

482
00:33:42,040 --> 00:33:44,440
You've just got the demo videos for now.

483
00:33:44,440 --> 00:33:51,680
And there is some capabilities that they didn't feature in the live stream when they announced

484
00:33:51,680 --> 00:33:52,680
it.

485
00:33:52,680 --> 00:33:57,160
But when you go into the article where they talk about the capabilities of this model,

486
00:33:57,160 --> 00:34:00,440
they shared some of the image output.

487
00:34:00,440 --> 00:34:07,340
So at the moment if you ask for an image in Chatchipity it uses the Dali 3 model.

488
00:34:07,340 --> 00:34:15,000
But this model has baked into it image capabilities that it can output and there's some really

489
00:34:15,000 --> 00:34:16,000
cool examples.

490
00:34:16,000 --> 00:34:22,280
There's an example where there's a coaster and it's a coaster with like a marble top

491
00:34:22,280 --> 00:34:25,560
half and a wooden bottom half.

492
00:34:25,560 --> 00:34:31,320
And then they upload a photo of this coaster and they upload a separate photo of the OpenAI

493
00:34:31,320 --> 00:34:33,360
logo.

494
00:34:33,360 --> 00:34:38,920
And they say we want to mock up what the OpenAI logo would look like if we got it manufactured

495
00:34:38,920 --> 00:34:43,760
and printed or kind of engraved into the coaster.

496
00:34:43,760 --> 00:34:46,680
And the model does it.

497
00:34:46,680 --> 00:34:53,160
But it creates this image using, well it kind of merges them both together and it looks

498
00:34:53,160 --> 00:34:54,740
legit.

499
00:34:54,740 --> 00:34:59,400
It doesn't look like a diffusion model where you get something like if you've done with

500
00:34:59,400 --> 00:35:03,680
Dali, you might describe that scene to it and then it gives you an output where it's

501
00:35:03,680 --> 00:35:08,600
like oh yeah that's kind of what I was after but if you asked it to do it again it would

502
00:35:08,600 --> 00:35:11,200
be a slightly different version, it would be slightly tweaked.

503
00:35:11,200 --> 00:35:18,000
No this does look like the original image with the logo attached to it.

504
00:35:18,000 --> 00:35:24,040
Writing on images, so text on images, wildly improved.

505
00:35:24,040 --> 00:35:31,600
There was an image shared by Greg Brockman on Twitter or on X where it showed an image

506
00:35:31,600 --> 00:35:35,680
that it created where it was a photo of somebody writing on a blackboard and it was a full

507
00:35:35,680 --> 00:35:44,840
passage of text and it was the full text, legible, not with that weird almost hieroglyphics

508
00:35:44,840 --> 00:35:51,480
kind of text that you often get with the diffusion models.

509
00:35:51,480 --> 00:35:57,200
And all of that's because it's a full, multimodal text image, everything's in the same neural

510
00:35:57,200 --> 00:36:02,120
network and it can do all of these things just better.

511
00:36:02,120 --> 00:36:06,020
I need to get in and play with some of those image capabilities because I haven't had much

512
00:36:06,020 --> 00:36:09,920
of a go yet with that part of it if I'm completely honest.

513
00:36:09,920 --> 00:36:10,920
It's not launched.

514
00:36:10,920 --> 00:36:12,320
Oh we can't access it yet.

515
00:36:12,320 --> 00:36:17,120
No much like the voice, yeah not there yet.

516
00:36:17,120 --> 00:36:21,340
This is all just in the articles and on the previews.

517
00:36:21,340 --> 00:36:27,560
There's been some complaints as well that people who've got GPT set up, their GPTs can

518
00:36:27,560 --> 00:36:34,620
only run off of GPT-4 not 4.0 as it stands even though there's a lot of excitement obviously

519
00:36:34,620 --> 00:36:36,840
about using 4.0 for this.

520
00:36:36,840 --> 00:36:42,600
So yeah, I mean there's one more thing we'll talk about when it comes to chat GPT improvements

521
00:36:42,600 --> 00:36:48,180
in a minute but it's part of this, there's such a desperation I think with these tech

522
00:36:48,180 --> 00:36:51,800
companies now to be seen to be ahead that they'll talk to you about cool stuff that

523
00:36:51,800 --> 00:36:57,480
they can do and then you might get it in two or four or eight or twelve weeks or maybe

524
00:36:57,480 --> 00:37:00,760
never and that is a little bit frustrating for the user.

525
00:37:00,760 --> 00:37:06,760
Now to give open AI the due, there's a bunch of stuff 4.0 related that was available.

526
00:37:06,760 --> 00:37:13,040
Basically I had 4.0 on my phone about an hour after the live stream ended so props to them

527
00:37:13,040 --> 00:37:16,760
for that but of course there's some cool stuff that we can't access and that's kind of frustrating

528
00:37:16,760 --> 00:37:22,440
because we want to play with it and of course open AI organized their event for the Monday

529
00:37:22,440 --> 00:37:28,800
before Google I.O. on the Tuesday and as we'll talk about, live streaming video and real

530
00:37:28,800 --> 00:37:32,880
time responsive voice assistance was kind of part of what Google presented as well.

531
00:37:32,880 --> 00:37:38,560
So there's definitely an arms race here and open AI doing their classic trolling of having

532
00:37:38,560 --> 00:37:43,560
an event just before Google have an event to try and assert them in that by doing that.

533
00:37:43,560 --> 00:37:48,920
Did you see the tweet from Sam Altman about Google?

534
00:37:48,920 --> 00:37:51,320
Go on, tell the listeners.

535
00:37:51,320 --> 00:37:57,400
That's where he said, I try not to think about competitors too much but I cannot stop thinking

536
00:37:57,400 --> 00:38:01,720
about the aesthetic difference between open AI and Google and he's got two photos from

537
00:38:01,720 --> 00:38:08,920
their live streams, their respective live streams and it's the kind of mid-century quiet

538
00:38:08,920 --> 00:38:18,200
little open AI office stream on one and the big bold brash somewhat trippy Google live

539
00:38:18,200 --> 00:38:21,920
stream stage setup and the comparisons between the two.

540
00:38:21,920 --> 00:38:25,840
The thing that I thought was interesting in that was he's directly said, I try not to

541
00:38:25,840 --> 00:38:31,760
think about competitors too much and he's putting them side by side as competition now

542
00:38:31,760 --> 00:38:36,200
which I don't know that I've seen him do before.

543
00:38:36,200 --> 00:38:40,440
There's been some twitchiness on Twitter, stroke X actually.

544
00:38:40,440 --> 00:38:47,160
There was someone who used to be at open AI that's now at Google who shared a video of

545
00:38:47,160 --> 00:38:53,600
showcasing similar abilities like pre-Google I.O. to be honest as open AI just launched

546
00:38:53,600 --> 00:38:59,840
like an hour after the stream is like an attempt to sort of say, hey, we've got something similar

547
00:38:59,840 --> 00:39:04,720
tomorrow so don't think you're too cool and of course all these things are getting retweeted

548
00:39:04,720 --> 00:39:10,560
and there's loads of comments so it is really interesting to see how we've had social media

549
00:39:10,560 --> 00:39:15,400
for a long time and I don't know if this is just because it's something I'm super interested

550
00:39:15,400 --> 00:39:17,360
in.

551
00:39:17,360 --> 00:39:25,960
The brand awareness and PR work that the likes of Sam Altman are doing primarily using cryptic

552
00:39:25,960 --> 00:39:33,680
tweets as the main mechanism for driving ridiculous amounts of discussion on Twitter, on Reddit

553
00:39:33,680 --> 00:39:41,560
as the main mechanism of disseminating information, building hype like it's proper super modern

554
00:39:41,560 --> 00:39:47,800
PR 101 in terms of… to a certain extent they don't even really issue traditional

555
00:39:47,800 --> 00:39:48,800
press releases.

556
00:39:48,800 --> 00:39:52,000
They write blog posts, they write a tweet about it and then all the media picks it up

557
00:39:52,000 --> 00:39:56,440
off a blog post from a sort of marketing and comm standpoint.

558
00:39:56,440 --> 00:40:02,720
It's really quite interesting to watch and we had Sam Altman allude to a bunch of different

559
00:40:02,720 --> 00:40:06,320
stuff, a super cryptic in the run up to the event so yeah, I guess that's interesting

560
00:40:06,320 --> 00:40:11,320
but in the interest of time there's one more thing from OpenAI this week in that they've

561
00:40:11,320 --> 00:40:14,160
just improved their data analysis tool.

562
00:40:14,160 --> 00:40:20,340
So one of the key questions here is if I can have GPT 4.0 for free, why should I still

563
00:40:20,340 --> 00:40:24,000
pay for Plus and Teams and Enterprise and all this stuff?

564
00:40:24,000 --> 00:40:27,760
And to a certain extent that's still to be answered but one of the power features of

565
00:40:27,760 --> 00:40:33,680
current paid parts of ChatGPT is its data analysis tool which is really, really cool

566
00:40:33,680 --> 00:40:37,320
and I think it's today they have improved the tool.

567
00:40:37,320 --> 00:40:40,360
So now you can upload files directly from Google Drive and Microsoft OneDrive which

568
00:40:40,360 --> 00:40:41,440
we talked about earlier.

569
00:40:41,440 --> 00:40:46,280
You can interact with tables and charts in this new expandable view and you can customize

570
00:40:46,280 --> 00:40:50,980
and download charts for using your presentations and documents.

571
00:40:50,980 --> 00:40:55,440
There are also sort of improvements in its ability to understand data sets and perform

572
00:40:55,440 --> 00:41:03,760
analysis tasks and this is going to be available for people who are basically paid users.

573
00:41:03,760 --> 00:41:05,840
So it's very new news.

574
00:41:05,840 --> 00:41:08,280
We haven't had a chance to too much to play with it.

575
00:41:08,280 --> 00:41:12,360
I think it's been getting incrementally better over time and I haven't been talking about

576
00:41:12,360 --> 00:41:18,520
it to be honest because some of the analyses that I tried to run six months ago with GPT

577
00:41:18,520 --> 00:41:22,720
4.0 have got better with things like the release of GPT 4.0 Turbo where they haven't really

578
00:41:22,720 --> 00:41:23,720
talked about it.

579
00:41:23,720 --> 00:41:28,360
So I'm quite interested to get in and play with this because I do find it's quite useful

580
00:41:28,360 --> 00:41:34,440
for guiding me through analysis of data, suggesting analyses I might do and helping me understand

581
00:41:34,440 --> 00:41:37,480
some of the patterns in data a bit easier which is kind of interesting.

582
00:41:37,480 --> 00:41:40,520
I don't know what you think about this, Mian.

583
00:41:40,520 --> 00:41:44,480
I tried 4.0 to do some data analysis this week.

584
00:41:44,480 --> 00:41:50,000
I actually just got a dummy data set just to see what it would look like and it was

585
00:41:50,000 --> 00:41:54,960
like four different CSVs related to airline loyalty programs.

586
00:41:54,960 --> 00:41:58,760
It's a publicly available data set and I just chucked it in and said analyze this, visualize

587
00:41:58,760 --> 00:42:08,960
it and do a statistical analysis and it did a huge amount of output in a single shot.

588
00:42:08,960 --> 00:42:11,360
It does multiple steps of the analysis.

589
00:42:11,360 --> 00:42:16,800
It starts off telling you what data there is, just literally, okay, these are the files

590
00:42:16,800 --> 00:42:18,060
you've got.

591
00:42:18,060 --> 00:42:23,160
This is the kind of data I think we can get and it's a really, really long comprehensive

592
00:42:23,160 --> 00:42:29,040
analysis which is, I think it was one, two, three, four.

593
00:42:29,040 --> 00:42:31,720
There's multiple charts provided in this analysis.

594
00:42:31,720 --> 00:42:33,480
In fact, I've just counted them up.

595
00:42:33,480 --> 00:42:37,080
There's five charts that it provides in this single analysis.

596
00:42:37,080 --> 00:42:39,960
It gives you insight into what they mean.

597
00:42:39,960 --> 00:42:45,900
So it gives you seasonal trends, customer loyalty, high value segments and retention

598
00:42:45,900 --> 00:42:49,880
strategies to deploy or from just a single prompt.

599
00:42:49,880 --> 00:42:56,380
This is not a really detailed, highly configured prompt engineering piece.

600
00:42:56,380 --> 00:43:00,080
This is just like, hey, here's some data, analyze it and visualize it for me.

601
00:43:00,080 --> 00:43:02,080
Tell me what it means.

602
00:43:02,080 --> 00:43:05,340
Yeah, I'm definitely going to play with it next week.

603
00:43:05,340 --> 00:43:09,060
It's also, I don't think we've mentioned this yet, but it's also markedly faster.

604
00:43:09,060 --> 00:43:13,900
One of the things about GPT-4 is I could read its outputs faster than it could produce them,

605
00:43:13,900 --> 00:43:21,240
but now it just spills down the page much faster than I can read and I have to catch

606
00:43:21,240 --> 00:43:25,560
it up, which I kind of like because it means I can skim read a bit more, but it's also

607
00:43:25,560 --> 00:43:30,840
a slightly different experience if I'm honest because it's like, like if it's verbose,

608
00:43:30,840 --> 00:43:31,840
then you're like, oh, crumbs, man.

609
00:43:31,840 --> 00:43:35,560
I don't know if I could be bothered to read all of this, but when you're like word by

610
00:43:35,560 --> 00:43:36,800
word then you would.

611
00:43:36,800 --> 00:43:42,440
So I don't know if my sort of ability to pay attention to it is going to change somewhat.

612
00:43:42,440 --> 00:43:47,440
So loads of great stuff going on over at OpenAI and I think if you've only been using their

613
00:43:47,440 --> 00:43:52,640
free tool so far and you weren't very impressed with GPT 3.5, it's worth logging into your

614
00:43:52,640 --> 00:43:58,560
free account one more time and giving GPT-4.0 a try because I think you'll find it's quite

615
00:43:58,560 --> 00:44:01,240
a bit better than you remember.

616
00:44:01,240 --> 00:44:04,960
The other thing we recommend is go and read the launch blog post about this new model

617
00:44:04,960 --> 00:44:09,500
because there's lots of different nuances to sort of how it works and that we haven't

618
00:44:09,500 --> 00:44:16,360
had a chance to cover today such as it can apparently detect emotion in your voice.

619
00:44:16,360 --> 00:44:21,520
The tool can sing if you ask it to and there's a bunch of interesting videos where it's messing

620
00:44:21,520 --> 00:44:26,120
stuff up, which just goes to show it's still prone to errors and hallucinations and all

621
00:44:26,120 --> 00:44:29,280
the good stuff that we've had to come to expect with models.

622
00:44:29,280 --> 00:44:36,560
But yes, well done OpenAI for an interesting release and now hot on the heels of that,

623
00:44:36,560 --> 00:44:41,360
we've got Google I.O. Mime, what happened over at Google I.O.?

624
00:44:41,360 --> 00:44:45,760
What didn't happen and I think we might need another hour long episode just to get through

625
00:44:45,760 --> 00:44:46,760
it all.

626
00:44:46,760 --> 00:44:51,320
In fact, if you really want to get through everything, they have put together a blog

627
00:44:51,320 --> 00:44:57,280
with the 100 things announced at Google I.O. and I'm sure there was probably more than

628
00:44:57,280 --> 00:45:03,320
100 but they rounded it down just to get in.

629
00:45:03,320 --> 00:45:10,440
It really is a massive list and we could go down the rabbit hole of Android updates and

630
00:45:10,440 --> 00:45:15,960
developer updates but I think what we'll try and do, you know, keep it business and marketing

631
00:45:15,960 --> 00:45:21,800
focused otherwise, well, we're not going out here before Monday at this rate.

632
00:45:21,800 --> 00:45:23,720
Yeah, I agree.

633
00:45:23,720 --> 00:45:27,720
So the main things, we'll start with the model updates.

634
00:45:27,720 --> 00:45:35,640
There were some updates to Gemini 1.5 Pro that we have seen.

635
00:45:35,640 --> 00:45:40,800
This is now available to everybody and had been made available through Vertex and AI

636
00:45:40,800 --> 00:45:42,200
Studio a few weeks ago.

637
00:45:42,200 --> 00:45:51,440
But the main difference on 1.5 Pro is that it has now got another increase in its context

638
00:45:51,440 --> 00:45:52,560
window.

639
00:45:52,560 --> 00:45:59,240
It was already 1 million tokens in the context window and they've just doubled that.

640
00:45:59,240 --> 00:46:06,560
It's now a 2 million context window large language model which is immediately just kind

641
00:46:06,560 --> 00:46:07,860
of mind blowing.

642
00:46:07,860 --> 00:46:09,600
It's getting very doctor evil now, Martin.

643
00:46:09,600 --> 00:46:14,500
I think it's like 1 million tokens.

644
00:46:14,500 --> 00:46:18,040
How can you even get enough information to ram into that?

645
00:46:18,040 --> 00:46:20,880
I don't know but I'm sure we'll find use cases.

646
00:46:20,880 --> 00:46:27,600
No, and every inference that you've run, I think at a million tokens, it comes out at

647
00:46:27,600 --> 00:46:30,800
around $7 per input.

648
00:46:30,800 --> 00:46:31,800
Wow.

649
00:46:31,800 --> 00:46:34,840
So every time you input, $7 gone.

650
00:46:34,840 --> 00:46:36,880
Well, there and thereabouts.

651
00:46:36,880 --> 00:46:39,760
So yeah, that's pretty impressive.

652
00:46:39,760 --> 00:46:49,120
Now we haven't compared 1.5 Pro with Advanced 1.0.

653
00:46:49,120 --> 00:46:53,440
But that massive context window does give amazing capabilities in terms of recall and

654
00:46:53,440 --> 00:46:55,920
just enhancing its overall knowledge.

655
00:46:55,920 --> 00:47:03,180
They also announced a new Gemini 1.5 model which is 1.5 Flash which is a faster, efficient

656
00:47:03,180 --> 00:47:09,620
model designed for large scale applications where you don't need that additional capability

657
00:47:09,620 --> 00:47:13,640
and speed of inference is more important.

658
00:47:13,640 --> 00:47:17,280
That's also 2 million token context window.

659
00:47:17,280 --> 00:47:19,280
I've been playing with that one actually, Martin.

660
00:47:19,280 --> 00:47:21,040
I've been very lucky.

661
00:47:21,040 --> 00:47:29,920
Both OpenAI's new 4.0 model and Gemini 1.5 Pro 2 million context and Flash were available

662
00:47:29,920 --> 00:47:32,300
in my favorite tool, Magi, as you know.

663
00:47:32,300 --> 00:47:33,840
So I've gotten and played with that.

664
00:47:33,840 --> 00:47:37,480
One of the cool things about Flash is it's cheaper but it's still pretty good at picking

665
00:47:37,480 --> 00:47:42,000
out information from large bits of context like a cool transcript.

666
00:47:42,000 --> 00:47:47,440
So giving the same sort of power but with lower cost.

667
00:47:47,440 --> 00:47:51,640
I'm a bit reticent to push Magi too hard, like you said, because it costs a bajillion

668
00:47:51,640 --> 00:47:55,680
dollars to run too many large prompts through it.

669
00:47:55,680 --> 00:47:59,620
The way that Magi is structured is after a certain amount of context you start to pay

670
00:47:59,620 --> 00:48:00,620
for that.

671
00:48:00,620 --> 00:48:05,360
But early testing is, I definitely think it's got some interesting use cases as a faster,

672
00:48:05,360 --> 00:48:06,360
cheaper model.

673
00:48:06,360 --> 00:48:10,480
I think it's also interesting that it's just so hard to keep up with all these model names

674
00:48:10,480 --> 00:48:11,600
and what they do.

675
00:48:11,600 --> 00:48:16,400
It's like, this is not like product development, it's like GitHub forking because they're kind

676
00:48:16,400 --> 00:48:21,280
of sort of pushed advanced into the background a bit now.

677
00:48:21,280 --> 00:48:27,600
And I think if you are a paid workspace user, a Google workspace user, my impression, I'd

678
00:48:27,600 --> 00:48:33,320
have to validate this, but my impression is we're on advance now but the 1.5 Pro's large

679
00:48:33,320 --> 00:48:36,840
context window and some of the interesting things it can do is going to be the model

680
00:48:36,840 --> 00:48:43,520
that gets baked into work place instead of Gemini Advanced.

681
00:48:43,520 --> 00:48:44,520
Super confusing.

682
00:48:44,520 --> 00:48:46,760
So, who knows what model we've got?

683
00:48:46,760 --> 00:48:51,440
Well, Gemini Advanced now has a 1 million token context window.

684
00:48:51,440 --> 00:48:52,440
Oh, okay.

685
00:48:52,440 --> 00:48:57,440
So they are trying to, they're running them concordantly, like what's the, I don't know.

686
00:48:57,440 --> 00:49:01,640
And if 1 million isn't enough for you Paul, then crikey.

687
00:49:01,640 --> 00:49:07,960
Dude, I want to feed it every film I've ever seen that I love and tell it to create a new

688
00:49:07,960 --> 00:49:09,960
movie for me on the fly.

689
00:49:09,960 --> 00:49:10,960
We're not there yet.

690
00:49:10,960 --> 00:49:14,800
We're not there yet people, but we could be soon.

691
00:49:14,800 --> 00:49:16,600
The Gemini Advanced updates were quite interesting.

692
00:49:16,600 --> 00:49:19,200
So they've expanded the context window.

693
00:49:19,200 --> 00:49:25,880
So that means you can analyze 1500 pages of PDF document.

694
00:49:25,880 --> 00:49:29,880
You can upload directly from Google Drive.

695
00:49:29,880 --> 00:49:36,400
They also include a data analysis and chart building for updated data files.

696
00:49:36,400 --> 00:49:38,400
And they've added this.

697
00:49:38,400 --> 00:49:43,960
So this is more of a consumer facing functionality, but they've added custom itinerary creation

698
00:49:43,960 --> 00:49:45,540
for travel planning.

699
00:49:45,540 --> 00:49:51,560
And it looks at the keywords are used in a different way and is able to build an itinerary

700
00:49:51,560 --> 00:49:52,880
for your trip.

701
00:49:52,880 --> 00:49:59,760
Now that I've already tried this out with things like ChatGPT for my own little travels

702
00:49:59,760 --> 00:50:04,680
and I've always found it to be pretty good at finding some places that I might not have

703
00:50:04,680 --> 00:50:06,280
thought about before.

704
00:50:06,280 --> 00:50:09,560
But yeah, this is this has got this as well now.

705
00:50:09,560 --> 00:50:16,720
They've got Gemini Live with 7 ounce, which is a natural intuitive spoken conversation,

706
00:50:16,720 --> 00:50:21,960
which apparently can be used for customer service and support similar to 4.0.

707
00:50:21,960 --> 00:50:22,960
I think this will be important.

708
00:50:22,960 --> 00:50:29,400
I think this would be important, Martin, because they're calling it like Project Astra and

709
00:50:29,400 --> 00:50:33,480
you can you can basically share a real time feed of what's going on around you and have

710
00:50:33,480 --> 00:50:38,400
a conversation with it very much like the sort of Omni channel GPT 4.0.

711
00:50:38,400 --> 00:50:45,360
So the cynic in me is that OpenAI found out what Google were going to launch on the Tuesday

712
00:50:45,360 --> 00:50:48,760
and they had something similar and they thought, well, we better get that out on the Monday

713
00:50:48,760 --> 00:50:50,600
then, hadn't we?

714
00:50:50,600 --> 00:50:53,920
Because they seem pretty comparable to me.

715
00:50:53,920 --> 00:50:58,800
There's a slightly longer delay on Gemini's response time in terms of having a natural

716
00:50:58,800 --> 00:51:01,680
conversation, but you can do the same things you can with GPT 4.0.

717
00:51:01,680 --> 00:51:04,160
You can interrupt it mid flow.

718
00:51:04,160 --> 00:51:07,880
It's got much more emotive capability in its voice.

719
00:51:07,880 --> 00:51:11,520
How well it understands human emotions and tonality is not clear at the moment.

720
00:51:11,520 --> 00:51:18,320
But but yeah, it's it feels a bit much for muchness, which just goes to show within like

721
00:51:18,320 --> 00:51:22,120
a few days we're already like, OK, that's the state of AI now, like showing the next

722
00:51:22,120 --> 00:51:23,120
thing.

723
00:51:23,120 --> 00:51:24,120
Absolutely.

724
00:51:24,120 --> 00:51:25,840
It becomes stable stakes.

725
00:51:25,840 --> 00:51:27,320
Everybody's now looking at AnthropiCurrent.

726
00:51:27,320 --> 00:51:28,680
Come on, where's your voice assist?

727
00:51:28,680 --> 00:51:29,680
Right.

728
00:51:29,680 --> 00:51:35,440
It's mad like that human's ability to like not be impressed by this stuff anymore is

729
00:51:35,440 --> 00:51:36,440
kind of astounding.

730
00:51:36,440 --> 00:51:40,620
And it fits very much with what, you know, OpenAI's goal was, as you said at the beginning

731
00:51:40,620 --> 00:51:44,640
of the podcast, to release things incrementally so that we didn't have a big shock.

732
00:51:44,640 --> 00:51:49,240
And I think one of the byproducts of that is we actually adapt quite quickly and we're

733
00:51:49,240 --> 00:51:53,640
already saying what's next rather than being like, oh my gosh, the world is being completely

734
00:51:53,640 --> 00:51:58,640
turned upside down, which, you know, if you'd been living in a bunker for 18 months is exactly

735
00:51:58,640 --> 00:51:59,640
how you'd feel.

736
00:51:59,640 --> 00:52:02,800
You'd come out and you'd go, sorry, when did Star Trek get here?

737
00:52:02,800 --> 00:52:10,680
Like how is or how have we gone from one paragraph out from GPT three to saw a video generation

738
00:52:10,680 --> 00:52:15,920
and basically a computer I can speak to that laughs and jokes back with me like it's mind

739
00:52:15,920 --> 00:52:16,920
blowing really.

740
00:52:16,920 --> 00:52:17,920
But yeah, that's happened.

741
00:52:17,920 --> 00:52:21,360
If you think about Jack GPT was November 2022.

742
00:52:21,360 --> 00:52:28,920
I think people that that was no time ago at all before that most people, I mean today,

743
00:52:28,920 --> 00:52:31,240
most people haven't really dug into the last language models.

744
00:52:31,240 --> 00:52:36,380
But before then, it was a tiny proportion of people that had ever used large language

745
00:52:36,380 --> 00:52:37,380
models.

746
00:52:37,380 --> 00:52:38,380
Yeah, it's amazing.

747
00:52:38,380 --> 00:52:42,320
So what else have we got coming out of the IO event?

748
00:52:42,320 --> 00:52:43,360
Image generation.

749
00:52:43,360 --> 00:52:50,000
They have updated their image generation diffusion model, Imagine, which is now Imagine 3, which

750
00:52:50,000 --> 00:52:54,720
is available to trusted testers on the Google platform.

751
00:52:54,720 --> 00:52:56,400
And that's in image effects.

752
00:52:56,400 --> 00:53:01,160
And it's coming to the vertex AI platform later in the summer.

753
00:53:01,160 --> 00:53:08,020
They also announced Vio, which is a kind of Sora competitor.

754
00:53:08,020 --> 00:53:15,600
So generating 1080p resolution videos in various styles suitable for marketing videos, stock

755
00:53:15,600 --> 00:53:19,160
video creation, content creation, and that kind of thing.

756
00:53:19,160 --> 00:53:21,680
Did you get a look at any of the clips of Vio?

757
00:53:21,680 --> 00:53:26,240
Yeah, so I've been looking at what the Twittersphere's had to say about the image generation and

758
00:53:26,240 --> 00:53:29,320
the video generation for those that have been able to play with it and also those that have

759
00:53:29,320 --> 00:53:30,780
just watched the demo.

760
00:53:30,780 --> 00:53:35,540
And it would appear that the image generation is much better, probably dually three level,

761
00:53:35,540 --> 00:53:36,840
but not mid journey level.

762
00:53:36,840 --> 00:53:40,320
So it's probably mid journey that's still just edging ahead in terms of the most powerful

763
00:53:40,320 --> 00:53:41,880
image generator.

764
00:53:41,880 --> 00:53:45,240
The video examples I've seen are impressive.

765
00:53:45,240 --> 00:53:49,400
They're certainly better than what we used to have coming out of Runway, but they're

766
00:53:49,400 --> 00:53:53,720
not as good as Sora at first viewing, I would suggest.

767
00:53:53,720 --> 00:53:59,320
So, but I think what it tells you is the magic that created Sora was not unique because this

768
00:53:59,320 --> 00:54:02,200
is pretty close, even if it's maybe slightly behind.

769
00:54:02,200 --> 00:54:04,680
Yeah, the ones I've seen look great.

770
00:54:04,680 --> 00:54:08,400
Again, with all of these tools, though, it's about getting into the hands of real users,

771
00:54:08,400 --> 00:54:09,400
isn't it?

772
00:54:09,400 --> 00:54:13,640
Because as we discussed with Balloonhead, it's okay showing the final output that's

773
00:54:13,640 --> 00:54:18,120
been edited and manipulated, but until you get it in the hands of users and people start

774
00:54:18,120 --> 00:54:23,480
prompting it and going, God, this is nearly impossible unless I use a thousand prompts

775
00:54:23,480 --> 00:54:25,240
to get the video I want.

776
00:54:25,240 --> 00:54:29,240
We won't actually know how good these tools are.

777
00:54:29,240 --> 00:54:35,840
Now I just want to focus on one area for the digital marketing listeners in our audience,

778
00:54:35,840 --> 00:54:38,640
and that is search.

779
00:54:38,640 --> 00:54:43,600
Because there have been some massive announcements on the search front.

780
00:54:43,600 --> 00:54:48,040
Google has been paying attention to the search generative experience.

781
00:54:48,040 --> 00:54:55,040
SGE will know that Google has been testing for some time, inserting AI generated responses

782
00:54:55,040 --> 00:55:00,320
at the top of search results, bumping down all of the organic search content.

783
00:55:00,320 --> 00:55:06,440
And this is now being rolled out in what Google is calling Google AI previews.

784
00:55:06,440 --> 00:55:12,480
This will be rolled out in America first and then across the wider world soon after.

785
00:55:12,480 --> 00:55:19,320
And it's fair to say people that pay a lot of attention to this online and in the Twittersphere

786
00:55:19,320 --> 00:55:23,760
are not impressed.

787
00:55:23,760 --> 00:55:29,640
This is pulling out some very strange results.

788
00:55:29,640 --> 00:55:35,140
There are examples where people are saying, look, this is virtually just pure plagiarism.

789
00:55:35,140 --> 00:55:40,800
And there is a general consensus amongst SEO professionals that, hey, look, we are heading

790
00:55:40,800 --> 00:55:49,280
towards Google zero click environment where Google passes through very, very little organic

791
00:55:49,280 --> 00:55:52,800
traffic to informational searches.

792
00:55:52,800 --> 00:55:58,440
Yeah, this is we've talked about this on the podcast multiple times.

793
00:55:58,440 --> 00:56:04,680
I think even when I think about my own behavior, thanks to you, I probably turn to perplexity

794
00:56:04,680 --> 00:56:08,360
as often as I do turn to Google, if not more often.

795
00:56:08,360 --> 00:56:09,360
Perplexity is great.

796
00:56:09,360 --> 00:56:12,200
It gives you a load of sources, it tells you where it got its insights.

797
00:56:12,200 --> 00:56:13,520
I don't click through to visit them.

798
00:56:13,520 --> 00:56:17,680
In fact, if I want more clarification, I ask a follow up question of perplexity.

799
00:56:17,680 --> 00:56:21,080
I don't go clicking through and reading stuff if I'm completely honest.

800
00:56:21,080 --> 00:56:25,840
The only time I'll ever click through is if my Spidey sense is like, no, I don't think

801
00:56:25,840 --> 00:56:26,840
that's right.

802
00:56:26,840 --> 00:56:30,080
And I'm basically validating hallucinations when I'm clicking through.

803
00:56:30,080 --> 00:56:36,960
So I think there's a use of behavior change coming here that's also in tandem with technology

804
00:56:36,960 --> 00:56:37,960
changes.

805
00:56:37,960 --> 00:56:45,120
And as you say, as a content publisher, not just traditional publishers, but basically

806
00:56:45,120 --> 00:56:50,280
companies leveraging content marketing to drive traffic through SEO and sell products

807
00:56:50,280 --> 00:56:57,080
and services, to what extent is that going to work well in X timeframe?

808
00:56:57,080 --> 00:57:03,320
And I think one of the hard things with this is to predict exactly how SEO shifts and that

809
00:57:03,320 --> 00:57:04,320
timeframe.

810
00:57:04,320 --> 00:57:10,560
Because it could take 12 months for SGE to just get really good and be rolled out widely.

811
00:57:10,560 --> 00:57:14,520
And then the amount of clicks that people are clicking through to informational content

812
00:57:14,520 --> 00:57:16,200
just drops off a cliff.

813
00:57:16,200 --> 00:57:20,880
Or the whole shift in behavior could be five years.

814
00:57:20,880 --> 00:57:27,700
If we're speaking with computers, certainly maybe ChatGBT or Gemini will pull up a source

815
00:57:27,700 --> 00:57:30,580
while we're having a conversation with it and say, hey, look, this is the article where

816
00:57:30,580 --> 00:57:32,120
I got some of these insights from.

817
00:57:32,120 --> 00:57:36,400
But if I'm not even looking at my screen, because I'm just getting the information from

818
00:57:36,400 --> 00:57:42,680
a direct conversation, this is going to cause fundamental shifts in terms of how people

819
00:57:42,680 --> 00:57:45,280
go to market, especially if they rely a lot on SEO.

820
00:57:45,280 --> 00:57:53,440
And we've talked a lot about how more than ever brand becomes extremely important, especially

821
00:57:53,440 --> 00:57:56,960
if we've already got an established brand, because I think establishing a new brand through

822
00:57:56,960 --> 00:58:01,760
content marketing is going to get harder and harder.

823
00:58:01,760 --> 00:58:06,280
And leveraging the expertise of your, if you're in a B2B environment, leveraging the expertise

824
00:58:06,280 --> 00:58:14,040
of your SMEs whose knowledge maybe doesn't exist on the web and therefore cannot be surfaced

825
00:58:14,040 --> 00:58:15,840
by an AI tool.

826
00:58:15,840 --> 00:58:21,040
And therefore your voice is still a trusted voice because you can share things that AI

827
00:58:21,040 --> 00:58:22,040
can't.

828
00:58:22,040 --> 00:58:26,240
So it's going to be really interesting to see this play out, isn't it, Mian?

829
00:58:26,240 --> 00:58:27,240
It is.

830
00:58:27,240 --> 00:58:32,360
So I think it's a chicken and egg in terms of how, maybe that's not the right phrase,

831
00:58:32,360 --> 00:58:36,720
chicken and egg, but it's a conundrum for publishers to think about how they want to

832
00:58:36,720 --> 00:58:37,720
treat the AIs.

833
00:58:37,720 --> 00:58:42,640
The AIs need the latest information.

834
00:58:42,640 --> 00:58:47,840
So if you're a publisher, you've got the researchers, you've got the expertise, the authority, all

835
00:58:47,840 --> 00:58:52,800
of the stuff that Google says you currently need, the EAT framework, you know, experience,

836
00:58:52,800 --> 00:58:55,080
expertise, authority, trustworthiness.

837
00:58:55,080 --> 00:58:59,040
You've got all of that within your industry and you can be writing about it.

838
00:58:59,040 --> 00:59:00,200
And it's the same with the press, right?

839
00:59:00,200 --> 00:59:04,760
If you've got, if you're the Washington Post or Financial Times, you've got journalists

840
00:59:04,760 --> 00:59:09,120
that are subject matter experts that go out and gather the information.

841
00:59:09,120 --> 00:59:14,040
And the AI then needs to troll that information, otherwise it won't know it.

842
00:59:14,040 --> 00:59:18,960
Now if Google is always serving, if people's behavior changes, that they just expect the

843
00:59:18,960 --> 00:59:25,760
AI to provide the answer, you as a publisher need to determine whether you want to feed

844
00:59:25,760 --> 00:59:26,760
that in.

845
00:59:26,760 --> 00:59:33,720
Now Google has made it available to, or have made the functionality available to basically

846
00:59:33,720 --> 00:59:35,400
de-index from search.

847
00:59:35,400 --> 00:59:40,680
So you can say, index me for search, but don't index me for the AI search specifically.

848
00:59:40,680 --> 00:59:45,080
So remove me from your previews tool.

849
00:59:45,080 --> 00:59:47,080
And we've seen a similar thing with OpenAI.

850
00:59:47,080 --> 00:59:50,440
They've made the same thing available as well.

851
00:59:50,440 --> 00:59:56,440
But then you're potentially, your business is not getting mentioned in all of these things

852
00:59:56,440 --> 01:00:02,600
where people are going to increasingly be turning to for their answers.

853
01:00:02,600 --> 01:00:03,780
So what do publishers do?

854
01:00:03,780 --> 01:00:08,840
Do you have to strike a deal with these companies like OpenAI has done recently, signing lots

855
01:00:08,840 --> 01:00:16,360
of deals with the likes of Reddit and large publishing companies are signing deals so

856
01:00:16,360 --> 01:00:19,520
that OpenAI can use their data?

857
01:00:19,520 --> 01:00:22,560
Is that how they start revenue generating?

858
01:00:22,560 --> 01:00:23,560
What about if you're a small B2B?

859
01:00:23,560 --> 01:00:29,120
You know, if you're a manufacturer who's done really well by positioning yourself as an

860
01:00:29,120 --> 01:00:33,760
expert in the field, but all of a sudden your blogs are being scraped by the AI and the

861
01:00:33,760 --> 01:00:40,120
AI is just going, yes, this is the answer about that really niche special thing that

862
01:00:40,120 --> 01:00:43,860
normally would have sent a load of traffic to your website for, but now we're just giving

863
01:00:43,860 --> 01:00:47,600
them the summary straight away on the search page.

864
01:00:47,600 --> 01:00:48,600
It's a tricky one.

865
01:00:48,600 --> 01:00:49,600
Yeah.

866
01:00:49,600 --> 01:00:54,440
And I think the publishers are probably doing the right thing.

867
01:00:54,440 --> 01:01:00,080
And I think honestly, we'll win if we have good licensing deals between the publishers

868
01:01:00,080 --> 01:01:06,280
and the AI models, because we need high quality journalism for a number of different reasons.

869
01:01:06,280 --> 01:01:10,080
And as you said, if the information is not there for AI's to scrape, then the AI's won't

870
01:01:10,080 --> 01:01:12,200
know the things that we need them to know.

871
01:01:12,200 --> 01:01:18,680
I think the challenge when you're a small or a mid-sized B2B or even a massive B2B company,

872
01:01:18,680 --> 01:01:24,440
to be honest, is your business model is not the information you produce.

873
01:01:24,440 --> 01:01:28,120
This has been like, content marketing has basically been an element of we give some

874
01:01:28,120 --> 01:01:32,120
of our expertise away as a freemium type model in the hope that you'll come buy some products

875
01:01:32,120 --> 01:01:35,480
and services from us because you'll trust us and think we're awesome.

876
01:01:35,480 --> 01:01:41,220
And does the commercial imperative hold up if you're producing content that's consumed

877
01:01:41,220 --> 01:01:45,320
by AI's that drives the advice and information they give but without citing you as a source?

878
01:01:45,320 --> 01:01:48,160
And ultimately, I think the answer to that probably is no.

879
01:01:48,160 --> 01:02:00,440
So either B2B has to move away from that type of model or OpenAI, Google find ways to either

880
01:02:00,440 --> 01:02:06,120
through paid means or organic means, better cite the sources that they're using.

881
01:02:06,120 --> 01:02:09,440
And I can see that paid means is the most likely.

882
01:02:09,440 --> 01:02:15,360
One of the key questions for large language models is, for OpenAI and Google, if Google

883
01:02:15,360 --> 01:02:20,480
disrupts its ad revenue, its pay-per-click Google ad revenue through this model, that's

884
01:02:20,480 --> 01:02:22,840
where it gets most of its cash.

885
01:02:22,840 --> 01:02:24,900
So it's not going to let that happen.

886
01:02:24,900 --> 01:02:28,760
So at some point, we have to expect to see some sort of advertising elements baked into

887
01:02:28,760 --> 01:02:29,760
this.

888
01:02:29,760 --> 01:02:37,320
The other thing is, are consumers going to consistently spend $20, $30, $40, $50 a month

889
01:02:37,320 --> 01:02:41,560
on a chat GPT or is it going to be mostly used by businesses?

890
01:02:41,560 --> 01:02:44,880
And this is again, a question we don't know.

891
01:02:44,880 --> 01:02:49,720
So if people are not willing to pay a monthly subscription for it, it will need to be an

892
01:02:49,720 --> 01:02:54,000
advertising driven model, which again, will completely influence the answers that are

893
01:02:54,000 --> 01:02:59,160
given, how you trust them, how companies, products and services are weaved into those.

894
01:02:59,160 --> 01:03:04,280
So I still think there's a bunch of unanswered questions here, but what advice would you

895
01:03:04,280 --> 01:03:10,720
give to a B2B company, Martin, that maybe produces lots of written content relying on

896
01:03:10,720 --> 01:03:11,720
search?

897
01:03:11,720 --> 01:03:15,680
Like, you know, what advice would we give them at this point in time, given potential

898
01:03:15,680 --> 01:03:18,000
avenues that this could go down?

899
01:03:18,000 --> 01:03:23,000
New formats, new channels, video, audio podcasts.

900
01:03:23,000 --> 01:03:27,660
I would make sure that you are turning that content into different channels.

901
01:03:27,660 --> 01:03:32,280
They still have legs, you know, they still has reach and it can all be turned into written

902
01:03:32,280 --> 01:03:36,120
content and you can actually use the AI to your advantage by taking a transcript from

903
01:03:36,120 --> 01:03:39,080
a podcast and turning into a series of blogs.

904
01:03:39,080 --> 01:03:43,960
But yeah, I would be looking at that, you know, social and influence, influencer type

905
01:03:43,960 --> 01:03:47,560
content, even in B2B works very well.

906
01:03:47,560 --> 01:03:51,880
You know, it might not be that you're going to be out there making loads of TikToks, might

907
01:03:51,880 --> 01:03:59,280
be, but find your channel and stick to producing content regularly for that channel in multimodal.

908
01:03:59,280 --> 01:04:02,800
I don't think that blogs are going to be here forever.

909
01:04:02,800 --> 01:04:03,800
They might supplement.

910
01:04:03,800 --> 01:04:10,000
But yeah, I would be looking to diversify the range of media formats that you produce

911
01:04:10,000 --> 01:04:11,000
for.

912
01:04:11,000 --> 01:04:12,000
Yeah, I'd agree with that.

913
01:04:12,000 --> 01:04:19,480
And I also think the, it's almost old school now, but a lot of the original content marketing

914
01:04:19,480 --> 01:04:23,020
business models were about generating subscribers.

915
01:04:23,020 --> 01:04:25,680
And I think subscribers becomes critical, right?

916
01:04:25,680 --> 01:04:30,600
Because you can no longer rely potentially on search for your content to be found.

917
01:04:30,600 --> 01:04:32,320
You need to become a trusted voice.

918
01:04:32,320 --> 01:04:33,320
That's always been true.

919
01:04:33,320 --> 01:04:36,960
That's always underpinned a lot of content marketing, but you need people to subscribe

920
01:04:36,960 --> 01:04:41,320
to your YouTube channel, to your podcast, to your newsletter so that you have a means

921
01:04:41,320 --> 01:04:46,160
of getting in front of them consistently and being the trusted voice that they listen to.

922
01:04:46,160 --> 01:04:51,880
So I think you combine all of those things together and you can ride out the diminishing

923
01:04:51,880 --> 01:04:53,640
search traffic.

924
01:04:53,640 --> 01:05:01,280
But I think if you're relying on keyword optimizing blog posts as your only strategy, I think

925
01:05:01,280 --> 01:05:04,880
that probably doesn't end well over the next two to three years.

926
01:05:04,880 --> 01:05:09,200
And you know, best time to plant a tree is 10 years ago.

927
01:05:09,200 --> 01:05:10,920
Second best time to plant a tree is today.

928
01:05:10,920 --> 01:05:14,720
So this is, if you're not thinking about this, I think today is the time to start thinking

929
01:05:14,720 --> 01:05:15,720
about that.

930
01:05:15,720 --> 01:05:18,640
And of course you can always contact us because we think about this a lot and we may be able

931
01:05:18,640 --> 01:05:19,640
to help you.

932
01:05:19,640 --> 01:05:24,200
There were some other things that I really thought were interesting at the event, Martin,

933
01:05:24,200 --> 01:05:29,740
and it's kind of, I'm surprised it's flown a little bit under the radar, but they reference

934
01:05:29,740 --> 01:05:33,920
something called AI teammate that they've been playing with internally.

935
01:05:33,920 --> 01:05:39,760
And they kind of inferred that this is what we can expect to start seeing in 2025.

936
01:05:39,760 --> 01:05:46,880
And in essence, it's like someone in your chat, like where the offices is Google, so

937
01:05:46,880 --> 01:05:50,200
it was in Google chat that is part of the conversation.

938
01:05:50,200 --> 01:05:54,680
I thought, I think it was called chip or clip or something like that.

939
01:05:54,680 --> 01:05:56,280
And it was a member of the chat.

940
01:05:56,280 --> 01:05:59,520
So you had the project team members there, you had chip, if indeed that's what it was

941
01:05:59,520 --> 01:06:02,800
even called, conversing back and forth.

942
01:06:02,800 --> 01:06:07,320
And the human would say, where are we on project X and what are the bottlenecks that's holding

943
01:06:07,320 --> 01:06:08,520
us back?

944
01:06:08,520 --> 01:06:12,780
And then the AI teammate would go check all the chat messages, all the relevant emails,

945
01:06:12,780 --> 01:06:16,860
all the relevant documents saved on the Google Drive system, and then come back with an update

946
01:06:16,860 --> 01:06:20,880
on where the project's at and what the bottlenecks are and what decisions still need to be made

947
01:06:20,880 --> 01:06:22,160
and by who.

948
01:06:22,160 --> 01:06:26,320
And then multiple people start having this conversation where the AI teammate is very

949
01:06:26,320 --> 01:06:28,280
much part of that conversation.

950
01:06:28,280 --> 01:06:39,440
And it was just a really interesting demonstration of this agent type approach, but in an extremely

951
01:06:39,440 --> 01:06:42,200
accessible business context, right?

952
01:06:42,200 --> 01:06:45,620
Like just someone in your chat that could well be a human really.

953
01:06:45,620 --> 01:06:47,040
And that's kind of how the interface worked.

954
01:06:47,040 --> 01:06:48,040
I thought it was interesting.

955
01:06:48,040 --> 01:06:49,040
Did you see this, Martin?

956
01:06:49,040 --> 01:06:50,040
What did you think?

957
01:06:50,040 --> 01:06:54,520
Yeah, so it was the teammate chip and you have to basically tag it into information

958
01:06:54,520 --> 01:06:56,360
so you can copy it into email threads.

959
01:06:56,360 --> 01:07:01,640
And as long as it's given access to these files, if you're sharing files and project

960
01:07:01,640 --> 01:07:06,240
or like I say, copying it into email chains, it will have access to that information.

961
01:07:06,240 --> 01:07:13,400
So you do have to think of it as this extra teammate and then it will respond in the chat.

962
01:07:13,400 --> 01:07:22,400
I think the example that they gave with the image was on the video was Project Sapphire

963
01:07:22,400 --> 01:07:26,360
and somebody puts in the chat, are we on track for the launch?

964
01:07:26,360 --> 01:07:32,040
And then Chip jumps in and says, there are conflicting decisions on the target audience

965
01:07:32,040 --> 01:07:33,040
for launch.

966
01:07:33,040 --> 01:07:37,400
And then it's got references to documents where it talks about this conflict.

967
01:07:37,400 --> 01:07:41,340
It says, getting aligned on this is important for our marketing timelines.

968
01:07:41,340 --> 01:07:43,960
Here are the latest timeline for Project Sapphire.

969
01:07:43,960 --> 01:07:48,800
And then it lays out development timeline, design and prototyping, testing and marketing.

970
01:07:48,800 --> 01:07:55,440
It's all there right in the Google Chat interface, which I think that's super cool.

971
01:07:55,440 --> 01:08:01,280
And I think we can expect to see this kind of thing rolled out in the future with Teams

972
01:08:01,280 --> 01:08:10,120
and keep an eye on this because the Microsoft Developer Conference is coming up next week

973
01:08:10,120 --> 01:08:17,560
and Slack announced a new AI assistant.

974
01:08:17,560 --> 01:08:21,520
And if you're a Slack user, you might want to check the terms and conditions of that

975
01:08:21,520 --> 01:08:27,000
because all of your data is going into the training for that model.

976
01:08:27,000 --> 01:08:29,560
That's a conversation for another day.

977
01:08:29,560 --> 01:08:35,160
But yes, I think we can expect that these places where we're all working, chatting and

978
01:08:35,160 --> 01:08:41,560
collaborating are going to have these new AI assistants where we can tag them into email

979
01:08:41,560 --> 01:08:43,160
threads into documents.

980
01:08:43,160 --> 01:08:48,180
We'll be able to probably tag them on Word docs and into Excel sheets.

981
01:08:48,180 --> 01:08:53,600
And they'll just be part of the collaborative experience.

982
01:08:53,600 --> 01:08:55,520
Yeah, do you know, it's fascinating.

983
01:08:55,520 --> 01:09:01,000
We've talked before on the podcast about how we're on exponential curves in different areas

984
01:09:01,000 --> 01:09:06,440
and different technologies and different past discoveries, if you like, have enabled where

985
01:09:06,440 --> 01:09:08,120
we are today in strange ways.

986
01:09:08,120 --> 01:09:13,160
We've talked about how the advent of the internet and to a certain extent, even publishing and

987
01:09:13,160 --> 01:09:16,960
content marketing itself have been critical for creating the volume of content needed

988
01:09:16,960 --> 01:09:18,640
to train these models.

989
01:09:18,640 --> 01:09:23,180
Without the internet, you don't get this and without a bunch of content on pretty much

990
01:09:23,180 --> 01:09:27,760
every topic you can think of, these AI models have very limited knowledge.

991
01:09:27,760 --> 01:09:36,200
But when I reflect upon this type of tool, it at the moment somewhat relies on everything

992
01:09:36,200 --> 01:09:37,720
being digital, right?

993
01:09:37,720 --> 01:09:44,640
By definition, it can only update you on your query if there is an email or a document somewhere

994
01:09:44,640 --> 01:09:46,600
that has that information in.

995
01:09:46,600 --> 01:09:52,360
So in-person meetings that are not recorded or the audio is not captured in some form,

996
01:09:52,360 --> 01:09:54,800
it's going to be important information it doesn't have.

997
01:09:54,800 --> 01:10:01,080
But of course, post pandemic, most of us are doing a huge amount of what we do on Teams

998
01:10:01,080 --> 01:10:09,040
and other, you know, Google me and other platforms, Zoom and the use of Slack and Gmail and other

999
01:10:09,040 --> 01:10:13,000
tools maybe hasn't increased so much.

1000
01:10:13,000 --> 01:10:17,880
I mean, I can imagine the use of tools like Slack and messaging probably has, but without

1001
01:10:17,880 --> 01:10:22,160
the pandemic, I don't think you get this shift in user behavior to everything being online

1002
01:10:22,160 --> 01:10:27,320
that even gives these tools the ability to, you know, to be a meaningful part of those

1003
01:10:27,320 --> 01:10:28,320
conversations.

1004
01:10:28,320 --> 01:10:35,400
And for a lot of businesses that have encouraged fully back to the office working conditions,

1005
01:10:35,400 --> 01:10:40,000
maybe we're all going to have to wear pendants around our neck like the one that you've bought

1006
01:10:40,000 --> 01:10:43,560
to be able to capture the data to enable our tools to be informed.

1007
01:10:43,560 --> 01:10:50,160
Because if your chip assistant only has half the information, then to be honest, it becomes

1008
01:10:50,160 --> 01:10:53,160
fairly useless because it might just there might be something really important.

1009
01:10:53,160 --> 01:10:54,160
It doesn't know.

1010
01:10:54,160 --> 01:10:58,120
In-person meetings are the key here, aren't they?

1011
01:10:58,120 --> 01:11:01,720
Because I've started if I have an in-person meeting with a client, I've gone to their

1012
01:11:01,720 --> 01:11:05,040
office, I will say, can I record this?

1013
01:11:05,040 --> 01:11:06,600
I'll stick my phone on the table.

1014
01:11:06,600 --> 01:11:09,880
And, you know, I've never had anybody say no.

1015
01:11:09,880 --> 01:11:16,000
And I'll record the session, just capture the audience, capture everything that is said

1016
01:11:16,000 --> 01:11:17,780
in that room.

1017
01:11:17,780 --> 01:11:22,160
And then I just upload that to a Google Drive and I've already got that set up with a Zap

1018
01:11:22,160 --> 01:11:28,080
on Zapier, which does the transcription, turns in all of the action points and creates the

1019
01:11:28,080 --> 01:11:29,080
to do list.

1020
01:11:29,080 --> 01:11:33,320
And nine times out of ten, these are taking place at a client's office.

1021
01:11:33,320 --> 01:11:37,600
So by the time I drive back to the office, everything's there, everything's captured

1022
01:11:37,600 --> 01:11:39,440
and it's shared in a Google Doc.

1023
01:11:39,440 --> 01:11:43,840
And I've got my follow up email drafted for me ready to just tweak and make sure I've

1024
01:11:43,840 --> 01:11:46,600
got all the details done.

1025
01:11:46,600 --> 01:11:48,560
But that's going to become more commonplace.

1026
01:11:48,560 --> 01:11:53,040
And I think things like, you'll probably see specific hardware for it, right?

1027
01:11:53,040 --> 01:12:01,080
So the intercom or the meeting room technology that exists at the moment with all of the

1028
01:12:01,080 --> 01:12:05,200
cameras set up and all of that kind of stuff, that's going to have it baked in, isn't it?

1029
01:12:05,200 --> 01:12:06,740
Yeah, I think you're right.

1030
01:12:06,740 --> 01:12:11,540
Because the value of the data to enable these tools will be too important to not do it.

1031
01:12:11,540 --> 01:12:18,280
It will open up the usual privacy, confidentiality, data security issues that we're used to dealing

1032
01:12:18,280 --> 01:12:19,280
with.

1033
01:12:19,280 --> 01:12:23,760
But the productivity gains are likely to drive the behavior changes anyway, I would have

1034
01:12:23,760 --> 01:12:24,760
thought.

1035
01:12:24,760 --> 01:12:26,840
But we'll have to see how that plays out.

1036
01:12:26,840 --> 01:12:31,440
I think that's pretty much it from Google I.O. that we would talk about from a business

1037
01:12:31,440 --> 01:12:32,440
perspective.

1038
01:12:32,440 --> 01:12:36,520
As Martin said, there are over a hundred different announcements, especially like consumer app

1039
01:12:36,520 --> 01:12:40,400
driven things like Google Photos is getting a bunch of upgrades and things that you can

1040
01:12:40,400 --> 01:12:41,400
do with it.

1041
01:12:41,400 --> 01:12:42,840
So we're interested in that.

1042
01:12:42,840 --> 01:12:46,040
We do suggest going and reading the full blog post.

1043
01:12:46,040 --> 01:12:52,840
There was one thing that I saw someone post on LinkedIn that I think is valuable for marketers.

1044
01:12:52,840 --> 01:12:59,760
Chris Penn, who many listeners may know from the digital marketing world, he posted something

1045
01:12:59,760 --> 01:13:07,800
on LinkedIn observing that Google Chrome in the next update is going to have the smallest

1046
01:13:07,800 --> 01:13:13,520
Google Gemini model, which is Nano, baked into the browser.

1047
01:13:13,520 --> 01:13:18,620
So it's going to run locally on device in the browser.

1048
01:13:18,620 --> 01:13:27,080
And he speculated that you will be able to use Nano in any window that has a text drafting

1049
01:13:27,080 --> 01:13:28,080
element.

1050
01:13:28,080 --> 01:13:32,740
So whether you're doing an email in Gmail or writing a comment on Facebook, anywhere

1051
01:13:32,740 --> 01:13:38,820
that you can input text, he estimates that you will likely have access to Nano.

1052
01:13:38,820 --> 01:13:47,100
And if this is the case, you can expect a significant uplift in the amount of spam comments

1053
01:13:47,100 --> 01:13:53,640
if you run a Facebook page and LinkedIn page or any other community platform for your brand.

1054
01:13:53,640 --> 01:13:59,400
Yeah, I'm so glad you mentioned that because it is interesting it runs in browser.

1055
01:13:59,400 --> 01:14:04,320
People are sensitive about sending data off to these companies and if it runs in browser,

1056
01:14:04,320 --> 01:14:09,520
in theory, your data is more controllable, although I'd put an asterisk and a massive

1057
01:14:09,520 --> 01:14:16,880
question mark that was like 50 times the font size of the original text because I'm really

1058
01:14:16,880 --> 01:14:19,000
not sure that you can truly trust that.

1059
01:14:19,000 --> 01:14:25,280
But how long before those tools can start taking actions in the browser as well?

1060
01:14:25,280 --> 01:14:29,960
Being to a certain extent already does a little bit of what you just described in terms of

1061
01:14:29,960 --> 01:14:35,320
you can select text in like a Google Doc and then being open in the sidebar can help you

1062
01:14:35,320 --> 01:14:38,240
rewrite and things like that, which is interesting.

1063
01:14:38,240 --> 01:14:44,260
The other thing that we didn't mention was that Gemini is really starting to get much

1064
01:14:44,260 --> 01:14:45,960
more power in Gmail.

1065
01:14:45,960 --> 01:14:53,440
And I think the examples from Google I.O. were consumer focused, things like summarizing

1066
01:14:53,440 --> 01:15:01,360
emails around trying to figure out what roof to hire, I think, for a domestic requirement.

1067
01:15:01,360 --> 01:15:06,120
But of course, you could see how easily that could be applied to business.

1068
01:15:06,120 --> 01:15:13,040
So ultimately, there's going to be improvements in the ability of Gmail for you to search

1069
01:15:13,040 --> 01:15:18,520
for information, for you to summarize information, for you to automatically draft emails and

1070
01:15:18,520 --> 01:15:24,880
have Gemini basically take action on certain emails for more complex tasks so that you

1071
01:15:24,880 --> 01:15:27,360
can do a little bit less of the manual work.

1072
01:15:27,360 --> 01:15:32,120
So rather than having to read through an email thread or emails from lots of different suppliers,

1073
01:15:32,120 --> 01:15:36,520
for example, being able to pull that information out automatically.

1074
01:15:36,520 --> 01:15:42,400
And I think one of the other examples they showed was pulling out like 50 receipts from

1075
01:15:42,400 --> 01:15:43,400
emails.

1076
01:15:43,400 --> 01:15:50,800
And like, for example, if you're a sole practitioner running your own business and then basically

1077
01:15:50,800 --> 01:15:55,880
having it pull them out and add all of those receipts to a spreadsheet with a link to the

1078
01:15:55,880 --> 01:15:59,520
original PDF of the receipt, that was pretty cool.

1079
01:15:59,520 --> 01:16:01,440
Yeah, you're nodding, Martin.

1080
01:16:01,440 --> 01:16:02,440
Did you see that as well?

1081
01:16:02,440 --> 01:16:03,440
Yeah, yeah, absolutely.

1082
01:16:03,440 --> 01:16:10,000
And that as somebody who is constantly getting emails, it's like, have you got this receipt?

1083
01:16:10,000 --> 01:16:14,040
Have you got that receipt, that absolutely game changer for me.

1084
01:16:14,040 --> 01:16:16,520
I saw that and thought I need that in my life immediately.

1085
01:16:16,520 --> 01:16:17,520
Right.

1086
01:16:17,520 --> 01:16:25,480
But it's the gateway, right, to pulling contextually relevant information into a format.

1087
01:16:25,480 --> 01:16:32,400
And I think there's so many different avenues of where AI will impact on the workplace and

1088
01:16:32,400 --> 01:16:35,120
how we work.

1089
01:16:35,120 --> 01:16:42,000
Starting from annoying, time consuming administrative tasks like this one through to potentially

1090
01:16:42,000 --> 01:16:47,960
quite creative brainstorming strategic tasks where the AI is acting as a brainstorm partner

1091
01:16:47,960 --> 01:16:48,960
and a coach.

1092
01:16:48,960 --> 01:16:55,720
And it's, I think, I don't see any single one of those being like, yes, that's like

1093
01:16:55,720 --> 01:17:00,600
10x the effectiveness of me or my business or whatever.

1094
01:17:00,600 --> 01:17:04,600
But I think when you start to add them all up, that's where you'll start to see where

1095
01:17:04,600 --> 01:17:06,280
the big impact is.

1096
01:17:06,280 --> 01:17:08,480
So there's pros and cons to that.

1097
01:17:08,480 --> 01:17:11,160
The pro is you can take advantage of the things you need.

1098
01:17:11,160 --> 01:17:13,600
You don't have to dabble in the things you don't need.

1099
01:17:13,600 --> 01:17:19,320
The con is I think maybe some businesses will feel a bit let down when they see that they

1100
01:17:19,320 --> 01:17:23,280
only get a slight incremental gain here and a nice incremental gain there.

1101
01:17:23,280 --> 01:17:28,240
They'd like to be able to switch AI on and double productivity.

1102
01:17:28,240 --> 01:17:30,080
But I just don't think it's going to work out like that.

1103
01:17:30,080 --> 01:17:35,200
I think it's going to be use cases, applications, opportunities and problems and then applying

1104
01:17:35,200 --> 01:17:37,280
AI in specific ways across those.

1105
01:17:37,280 --> 01:17:43,360
And so mid to large companies are probably going to have some sort of chief AI officer

1106
01:17:43,360 --> 01:17:47,360
or someone who's in charge of understanding what technology is available, what they can

1107
01:17:47,360 --> 01:17:51,840
do and how to apply them to different use cases and applications in a business to start

1108
01:17:51,840 --> 01:17:55,760
to realize some of those efficiency and creativity gains.

1109
01:17:55,760 --> 01:18:01,240
Because that's going to probably need to be someone or a team's job in order to deploy

1110
01:18:01,240 --> 01:18:08,280
across those different functions, but also to be in charge of the change management because

1111
01:18:08,280 --> 01:18:12,320
we're going to have to coach and train and bring humans through on the journey to help

1112
01:18:12,320 --> 01:18:13,760
them actually adopt the tools.

1113
01:18:13,760 --> 01:18:15,840
What are your thoughts?

1114
01:18:15,840 --> 01:18:16,840
Absolutely agree.

1115
01:18:16,840 --> 01:18:24,840
And if you think about any piece of software, so Microsoft Word, Microsoft Word has so many

1116
01:18:24,840 --> 01:18:25,840
functionalities.

1117
01:18:25,840 --> 01:18:29,000
In fact, actually PowerPoint is probably a better example.

1118
01:18:29,000 --> 01:18:36,280
Most people can use PowerPoint and throw together a slide deck, but scratch the surface and

1119
01:18:36,280 --> 01:18:41,360
lots of people say it's one of the best graphic design tools out there on the market.

1120
01:18:41,360 --> 01:18:46,280
The functionality that is kind of hidden away and what you can do with it when you scratch

1121
01:18:46,280 --> 01:18:49,400
the surface is remarkable.

1122
01:18:49,400 --> 01:18:52,400
And most people use about 10% of it.

1123
01:18:52,400 --> 01:18:57,880
Most Excel you considered to be an intermediate borderline advanced user.

1124
01:18:57,880 --> 01:19:02,520
If you can use pivot tables and V lookups.

1125
01:19:02,520 --> 01:19:04,980
These are fairly basic functionalities.

1126
01:19:04,980 --> 01:19:09,400
But if you're a real power user, there's a huge amount that you can do with it.

1127
01:19:09,400 --> 01:19:11,360
But most people don't use it.

1128
01:19:11,360 --> 01:19:18,360
And with AI features and functionalities being rolled out, like the AI extraction of attachments

1129
01:19:18,360 --> 01:19:24,920
and automatic organization of those for users, unless people know that they're there and

1130
01:19:24,920 --> 01:19:30,680
how to navigate them, they'll just sit there untapped.

1131
01:19:30,680 --> 01:19:37,120
Because as identified with the examples I just gave, most features and functionality

1132
01:19:37,120 --> 01:19:41,360
does sit there untapped by most ordinary users.

1133
01:19:41,360 --> 01:19:47,220
Yeah, I think it's going to be super interesting to see how AI enables that.

1134
01:19:47,220 --> 01:19:51,640
Because you also don't know what you don't know.

1135
01:19:51,640 --> 01:19:56,720
So there's probably cool stuff that I could do in Excel to ask interesting questions of

1136
01:19:56,720 --> 01:19:58,560
data and draw interesting insights.

1137
01:19:58,560 --> 01:20:01,440
But if I don't know what questions to ask, then I'm not even going to start to think

1138
01:20:01,440 --> 01:20:05,080
about how I might be able to actually achieve that in terms of analysis.

1139
01:20:05,080 --> 01:20:09,920
So again, AI as a coach going, you know, cool, what is it that you're actually looking to

1140
01:20:09,920 --> 01:20:10,920
find out here?

1141
01:20:10,920 --> 01:20:15,160
And maybe even having enough context about me, my business, the data I'm looking at

1142
01:20:15,160 --> 01:20:18,160
to go, do you know what you might find interesting is this?

1143
01:20:18,160 --> 01:20:20,360
And I'll go, yeah, that does sound interesting.

1144
01:20:20,360 --> 01:20:22,380
How could we go about analyzing that?

1145
01:20:22,380 --> 01:20:28,080
And then as we see with the ongoing improvements with the data analysis tool in chat GPT, the

1146
01:20:28,080 --> 01:20:34,440
actual ability to write and run Python code to undertake a series of analysis steps at

1147
01:20:34,440 --> 01:20:40,840
this point, let's be honest, to actually dive into that data and provide those insights,

1148
01:20:40,840 --> 01:20:43,320
I think that's going to be quite fun.

1149
01:20:43,320 --> 01:20:50,080
And I think overall, I think as I said earlier in the podcast, having that AI assistant is

1150
01:20:50,080 --> 01:20:56,400
going to drag up our skill levels in all the areas where we're a bit weaker and there's

1151
01:20:56,400 --> 01:20:59,280
lots of stuff that we don't know, because you're just going to have that AI on your

1152
01:20:59,280 --> 01:21:01,560
shoulder going, hey, can I help you with this?

1153
01:21:01,560 --> 01:21:02,560
Have you thought about this?

1154
01:21:02,560 --> 01:21:06,960
And I think that's going to be quite fun and pretty valuable as long as we deploy it in

1155
01:21:06,960 --> 01:21:08,560
a human first way.

1156
01:21:08,560 --> 01:21:12,760
Because I think until we get into a mechanism where every bit of information is digital,

1157
01:21:12,760 --> 01:21:17,920
I think there'll be things that humans know about their work and their clients and their

1158
01:21:17,920 --> 01:21:23,220
colleagues and that the AIs won't know and you will need the humans piloting and driving

1159
01:21:23,220 --> 01:21:28,400
a lot of this and it's just about augmenting them the best that we can.

1160
01:21:28,400 --> 01:21:30,240
Well there you have it folks.

1161
01:21:30,240 --> 01:21:32,880
We whistle-stopped through a bunch of stuff that we missed.

1162
01:21:32,880 --> 01:21:35,440
Sorry for being out and about.

1163
01:21:35,440 --> 01:21:37,600
We've talked a bit about GPT 4.0.

1164
01:21:37,600 --> 01:21:42,440
We've talked about Google I.O. and loads of cool stuff.

1165
01:21:42,440 --> 01:21:47,000
We're back in the mix now so you're going to expect to see us dropping into your favourite

1166
01:21:47,000 --> 01:21:52,320
podcasting programme on a frequent basis, probably every week or two, a little bit like

1167
01:21:52,320 --> 01:21:54,040
what we used to back in the day.

1168
01:21:54,040 --> 01:21:58,400
And if you have enjoyed this podcast, please interact with one of our posts on LinkedIn.

1169
01:21:58,400 --> 01:22:00,920
We'd love to know what you thought of different things in the podcast.

1170
01:22:00,920 --> 01:22:02,720
You can leave us a review.

1171
01:22:02,720 --> 01:22:03,720
We'd love that.

1172
01:22:03,720 --> 01:22:08,240
You can tell your friends, hey, the guys are back and they are, didn't even mention Derby

1173
01:22:08,240 --> 01:22:11,560
County yet, who got promoted which we really should.

1174
01:22:11,560 --> 01:22:15,520
Yeah, the next five minutes, we're in now.

1175
01:22:15,520 --> 01:22:16,520
What a day.

1176
01:22:16,520 --> 01:22:19,080
What a day for all that was fantastic.

1177
01:22:19,080 --> 01:22:23,080
Promoted back in the championship where we belong.

1178
01:22:23,080 --> 01:22:27,800
See we don't, we only talk about Derby County when they're losing because that entertains

1179
01:22:27,800 --> 01:22:28,800
me.

1180
01:22:28,800 --> 01:22:31,520
But when they're winning and they're doing well, that might entertain you, Ryan.

1181
01:22:31,520 --> 01:22:33,200
And we can't have that.

1182
01:22:33,200 --> 01:22:34,760
That's outside of time.

1183
01:22:34,760 --> 01:22:38,680
Yeah, well now we have a long rebuilding job ahead of us.

1184
01:22:38,680 --> 01:22:47,080
But maybe to pull it back to the world of AI, Derby can tap into the AI tactics tool

1185
01:22:47,080 --> 01:22:54,560
that was developed by Google DeepMind and launched alongside Liverpool, your team.

1186
01:22:54,560 --> 01:22:56,880
That's right for a better corner taking.

1187
01:22:56,880 --> 01:23:01,460
But at least we've done the listeners the good grace of putting all of this in aim football

1188
01:23:01,460 --> 01:23:03,960
stuff at the end of the podcast.

1189
01:23:03,960 --> 01:23:09,720
Okay, right, we will let you go.

1190
01:23:09,720 --> 01:23:11,160
It's been a super long episode.

1191
01:23:11,160 --> 01:23:14,760
If you've made it this far, we appreciate you sticking with us and we look forward to

1192
01:23:14,760 --> 01:23:19,860
keeping you up to date on all the things you need to know about AI to help you be more

1193
01:23:19,860 --> 01:23:23,920
effective in your role, run your business better, be a better marketer, et cetera, et

1194
01:23:23,920 --> 01:23:26,440
cetera, in our next episode.

1195
01:23:26,440 --> 01:23:29,800
If just one more thing, if there's anything you'd like us to cover in a podcast in the

1196
01:23:29,800 --> 01:23:34,160
future, email us at hello at artificiallyintelligentmarketing.com.

1197
01:23:34,160 --> 01:23:37,320
I think that sounds like a blooming good idea.

1198
01:23:37,320 --> 01:23:38,320
We'd love to hear from you.

1199
01:23:38,320 --> 01:23:40,240
Right, we'll sign out there, Martin.

1200
01:23:40,240 --> 01:23:42,240
Thanks very much for your time.

1201
01:23:42,240 --> 01:23:43,240
Bye.

1202
01:23:43,240 --> 01:23:48,760
Thank you for listening to Artificially Intelligent Marketing.

1203
01:23:48,760 --> 01:23:54,920
To stay on top of the latest trends, tips and tools in the world of marketing AI, be

1204
01:23:54,920 --> 01:23:56,580
sure to subscribe.

1205
01:23:56,580 --> 01:24:00,160
We look forward to seeing you again next week.