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

5
00:00:23,240 --> 00:00:28,280
Hello listeners, welcome to episode 37 of Artificially Intelligent Marketing.

6
00:00:28,280 --> 00:00:32,880
It's Paul Avery here, joined as always by the fantabulous Martin Broadhurst.

7
00:00:32,880 --> 00:00:35,840
Mark Tino, how are you today my friend?

8
00:00:35,840 --> 00:00:43,080
I'm living my best Friday with elves in the background for those who are watching the

9
00:00:43,080 --> 00:00:44,080
video.

10
00:00:44,080 --> 00:00:46,760
I'm counting down to Christmas.

11
00:00:46,760 --> 00:00:53,720
Wouldn't normally be this Christmas prepared, except having a toddler in my life these days

12
00:00:53,720 --> 00:00:57,760
has just brought the whole thing forward a little bit.

13
00:00:57,760 --> 00:00:59,800
You are having an Elf of a Time.

14
00:00:59,800 --> 00:01:01,320
Cool, that one doesn't quite work.

15
00:01:01,320 --> 00:01:04,240
We're not going to cut it though, we're going to make you listen to those types of jokes

16
00:01:04,240 --> 00:01:05,760
all the way through the episode.

17
00:01:05,760 --> 00:01:09,360
Yeah, well I'm glad you're enjoying it, I'm loving your background.

18
00:01:09,360 --> 00:01:13,680
So much so I'm probably going to edit my background in post so that when people see this they'll

19
00:01:13,680 --> 00:01:18,360
go, oh wow, Paul made the effort on his background as well, which in real time I hadn't, but

20
00:01:18,360 --> 00:01:19,360
they will never know that.

21
00:01:19,360 --> 00:01:22,720
I mean they will now, but otherwise they wouldn't have known.

22
00:01:22,720 --> 00:01:26,200
We got some cracking stories this week, haven't we Martin?

23
00:01:26,200 --> 00:01:30,240
In fact, before we get into our first story I've got to tell you, as you know I was at

24
00:01:30,240 --> 00:01:37,160
a conference last week, I was at the SAMPS annual US meeting in Boston and we were talking

25
00:01:37,160 --> 00:01:38,160
about AI.

26
00:01:38,160 --> 00:01:43,880
While I was talking I did a couple of short talks on different topics and then there was

27
00:01:43,880 --> 00:01:48,280
a panel and in the panel I slightly bashed Google.

28
00:01:48,280 --> 00:01:53,320
I was like, they're trying to catch up, they're a bit late to the party, yadda yadda yadda.

29
00:01:53,320 --> 00:01:59,160
I sat down and one of my good friends, Jeremiah Wurst, Jeremiah, little shout out to you,

30
00:01:59,160 --> 00:02:04,260
he turned to me, shook me his phone and he went, you know that Google have just launched

31
00:02:04,260 --> 00:02:07,320
Gemini, why are you speaking?

32
00:02:07,320 --> 00:02:13,280
So I am going to take personal responsibility for accelerating the release of Google Gemini

33
00:02:13,280 --> 00:02:18,400
by giving Google a hard time on a panel with 70 people in the room that they clearly had

34
00:02:18,400 --> 00:02:22,880
a representative at and said we need to get this out now because I'm fed up with the

35
00:02:22,880 --> 00:02:27,160
folks at Artificially Intelligent Marketing giving us all sorts of shit about it.

36
00:02:27,160 --> 00:02:34,400
So the share price dropping in real time based on your comments and they said release press,

37
00:02:34,400 --> 00:02:36,400
please press release now.

38
00:02:36,400 --> 00:02:42,000
But it was quite funny and it also did really serve to really highlight another point that

39
00:02:42,000 --> 00:02:47,200
we made again and again during the day which is this stuff is moving lightning fast.

40
00:02:47,200 --> 00:02:52,840
When I started my talk my comment was fair, by the end of my talk it wasn't fair and

41
00:02:52,840 --> 00:02:53,840
I don't know anymore.

42
00:02:53,840 --> 00:02:59,520
But that is AI in a nutshell and especially the applications and technologies we're seeing

43
00:02:59,520 --> 00:03:02,840
emerge for marketing and sales folks and business in general.

44
00:03:02,840 --> 00:03:08,440
So I think an overriding feeling at the event was it's extremely hard to stay on top of

45
00:03:08,440 --> 00:03:12,040
all of this stuff and if you're listening to this and that's your feeling, if you feel

46
00:03:12,040 --> 00:03:15,760
like you're drinking from a fire hose and you're completely overwhelmed, I don't think

47
00:03:15,760 --> 00:03:17,240
anybody feels any different.

48
00:03:17,240 --> 00:03:20,280
You are absolutely in the crew when it comes to that.

49
00:03:20,280 --> 00:03:26,320
Even nerds who spend many hours a week on this stuff like Martin and I will come together

50
00:03:26,320 --> 00:03:32,000
to do this podcast on a Friday and between us there'll be stories that Martin's seen

51
00:03:32,000 --> 00:03:36,320
and I haven't and vice versa because it's just so hard to stay on top of things.

52
00:03:36,320 --> 00:03:38,000
But that's why we give you this hour.

53
00:03:38,000 --> 00:03:41,720
Hopefully we can do at least 80% of the heavy lifting for you and with some of that heavy

54
00:03:41,720 --> 00:03:47,440
lifting in mind, Martin, why didn't you tell us about the biggest story this week?

55
00:03:47,440 --> 00:03:56,240
That Google company who said we're no good at AI, they've announced a GPT-4 competitor

56
00:03:56,240 --> 00:03:58,280
in the form of Gemini.

57
00:03:58,280 --> 00:04:02,920
So Gemini is their multimodal large language model.

58
00:04:02,920 --> 00:04:06,640
It actually comes in three sizes.

59
00:04:06,640 --> 00:04:14,920
So they've announced Nano, Thro and Ultra.

60
00:04:14,920 --> 00:04:22,280
Ultra is the big one, that's the one that is supposedly capable and well actually outperforming

61
00:04:22,280 --> 00:04:25,640
according to benchmarks GPT-3.

62
00:04:25,640 --> 00:04:31,880
I think it beats it on 30 out of 32 benchmark tests.

63
00:04:31,880 --> 00:04:36,120
Thro is the one that they've released now.

64
00:04:36,120 --> 00:04:42,280
It's effectively GPT-3.5.

65
00:04:42,280 --> 00:04:48,760
It's available via API and BARD if you're not in Europe or the UK.

66
00:04:48,760 --> 00:04:53,680
You can actually access it if you use a VPN, which I have done.

67
00:04:53,680 --> 00:04:59,000
I've played around with it and yeah, it's similar to GPT-3.5.

68
00:04:59,000 --> 00:05:01,200
It's fine.

69
00:05:01,200 --> 00:05:08,480
Ultra looks really interesting because it's multimodal in a way that I found quite interesting

70
00:05:08,480 --> 00:05:11,320
the way that they described it during the launch event.

71
00:05:11,320 --> 00:05:17,280
So Demis Hesabis, the CEO of DeepMind, they've clearly been working closely with Google Brain

72
00:05:17,280 --> 00:05:21,080
and DeepMind working together now.

73
00:05:21,080 --> 00:05:23,800
They've brought these two organizations together.

74
00:05:23,800 --> 00:05:32,400
Demis Hesabis said that this is a room multimodal model rather than an interface that gives

75
00:05:32,400 --> 00:05:34,600
the appearance of multimodality.

76
00:05:34,600 --> 00:05:41,760
So I think what he was referencing there was if you use ChatGPT, GPT-4 is multimodal in

77
00:05:41,760 --> 00:05:48,560
so much as you can input text and image, but it can only output text.

78
00:05:48,560 --> 00:05:54,760
But if you're using ChatGPT, you get the appearance of multimodality because it will output images

79
00:05:54,760 --> 00:06:01,000
through the model Dali, which is connected to it, and you can input voice.

80
00:06:01,000 --> 00:06:07,200
Again, that's actually using Whisper to transcribe it and then inputting it as text.

81
00:06:07,200 --> 00:06:09,000
And it can output voice.

82
00:06:09,000 --> 00:06:12,560
But again, that's not actually part of that same model.

83
00:06:12,560 --> 00:06:13,560
It's a different model.

84
00:06:13,560 --> 00:06:14,560
It's all connected up.

85
00:06:14,560 --> 00:06:20,120
Whereas Google is saying that Gemini is truly multimodal.

86
00:06:20,120 --> 00:06:26,640
And they actually give some interesting examples of this in the demo video.

87
00:06:26,640 --> 00:06:35,840
For example, they have somebody speaking in Chinese and Chinese dialect and asking it

88
00:06:35,840 --> 00:06:40,240
to translate into English.

89
00:06:40,240 --> 00:06:46,720
And they said this would actually not work in the current models that are on the market

90
00:06:46,720 --> 00:06:55,040
because the language, the input language is tonal and that would be lost in the Whisper

91
00:06:55,040 --> 00:06:57,560
transcription of it.

92
00:06:57,560 --> 00:07:04,800
So you would lose some of the nuance around the input of audio like that.

93
00:07:04,800 --> 00:07:06,720
It looks really powerful.

94
00:07:06,720 --> 00:07:11,740
The proof is in the pudding though, and that's what we're all waiting for to get our hands

95
00:07:11,740 --> 00:07:14,840
on it because it isn't available to users.

96
00:07:14,840 --> 00:07:21,120
Ultra isn't available to users until early 2024.

97
00:07:21,120 --> 00:07:22,320
So it looks very exciting.

98
00:07:22,320 --> 00:07:29,520
However, it was not launched without a little bit of controversy.

99
00:07:29,520 --> 00:07:38,040
The launch video showed it doing certain things that seemed very impressive, I must admit.

100
00:07:38,040 --> 00:07:46,800
So the example in the video was hand drawn images, little sketches on Post-it notes.

101
00:07:46,800 --> 00:07:51,760
For instance, two cars driving down a hill.

102
00:07:51,760 --> 00:07:56,280
One of them looking like an aerodynamic car, the other one like a block.

103
00:07:56,280 --> 00:08:00,600
And it says, which of these would go faster?

104
00:08:00,600 --> 00:08:07,080
And the model responds, the one on the right, which is more aerodynamic.

105
00:08:07,080 --> 00:08:10,640
And then they keep editing the pictures by drawing new bits on them and speaking to the

106
00:08:10,640 --> 00:08:14,120
model and it converses back and forth with them.

107
00:08:14,120 --> 00:08:18,080
And it looks like this whole thing takes place without any text input.

108
00:08:18,080 --> 00:08:22,880
It's just somebody speaking to the model, showing it pictures and changing the pictures

109
00:08:22,880 --> 00:08:23,880
and saying, what about this?

110
00:08:23,880 --> 00:08:25,040
What about that?

111
00:08:25,040 --> 00:08:26,040
And the model responding.

112
00:08:26,040 --> 00:08:35,000
Well, it turns out that whole segment was completely manipulated and not at all as described.

113
00:08:35,000 --> 00:08:40,240
The images were actually uploaded with text prompts.

114
00:08:40,240 --> 00:08:45,080
It was not a speech back and forth with images being shown in real time.

115
00:08:45,080 --> 00:08:48,480
The whole thing was a manipulation.

116
00:08:48,480 --> 00:08:56,240
And actually what people took umbrage with was that what the voice was saying as a kind

117
00:08:56,240 --> 00:09:01,520
of input prompt, so what the human was supposed to have been saying, didn't match what the

118
00:09:01,520 --> 00:09:05,040
text based prompt actually said.

119
00:09:05,040 --> 00:09:09,760
So they gave more detail in the actual text prompt that the model got.

120
00:09:09,760 --> 00:09:15,520
And people were quite annoyed about the manipulation there.

121
00:09:15,520 --> 00:09:20,320
But that aside, this model is very impressive on the face of it.

122
00:09:20,320 --> 00:09:22,840
It does have some very cool capabilities.

123
00:09:22,840 --> 00:09:28,920
At the other end of the scale, going from Gemini Ultra down to Gemini Nano, this is

124
00:09:28,920 --> 00:09:34,680
also very interesting because what they've announced here is a RNE model that you can

125
00:09:34,680 --> 00:09:41,920
actually run on a mobile phone without requiring any internet connection.

126
00:09:41,920 --> 00:09:48,580
And this is going to be rolled out to the Google Pixel 8 and the Google Pixel series

127
00:09:48,580 --> 00:09:49,580
in the future.

128
00:09:49,580 --> 00:09:59,120
So I'm very excited to see the evolution of these small language models for mobile devices.

129
00:09:59,120 --> 00:10:02,880
Yeah, there's a lot to unpack there.

130
00:10:02,880 --> 00:10:08,720
The tonality thing was very interesting and I imagined, and I would love it when people

131
00:10:08,720 --> 00:10:14,160
can get their hands on this, being able to speak with a bot where it's listening to your

132
00:10:14,160 --> 00:10:18,200
tone of voice and it can tell if you're stressed or scared or upset, right?

133
00:10:18,200 --> 00:10:24,360
So like we've talked before about the different applications of bots in the future and how

134
00:10:24,360 --> 00:10:32,040
when people like to rate doctor outputs as based on how empathetic they were, GPT-4 was

135
00:10:32,040 --> 00:10:34,840
rated more empathetic than real human doctors.

136
00:10:34,840 --> 00:10:38,800
Well, how much does that empathy increase when a model can actually understand the tone

137
00:10:38,800 --> 00:10:45,240
of voice, not just the text, like content of the information that's being presented?

138
00:10:45,240 --> 00:10:46,600
That's pretty interesting.

139
00:10:46,600 --> 00:10:48,120
On the...

140
00:10:48,120 --> 00:10:50,640
We fudged this to make it look better.

141
00:10:50,640 --> 00:10:51,920
What a shame.

142
00:10:51,920 --> 00:10:52,920
What a shame.

143
00:10:52,920 --> 00:11:00,200
So desperate to show they were back in the game that they just reached 10% too far.

144
00:11:00,200 --> 00:11:03,360
And because we know that, I think many of us are now...

145
00:11:03,360 --> 00:11:07,480
I think many people are like, all right, I'm not going to reserve my judgment until I see

146
00:11:07,480 --> 00:11:10,720
it in action, but you just lost my trust.

147
00:11:10,720 --> 00:11:17,640
So I'm personally, I'm assuming it's as good as GPT-4 and maybe slightly worse, honestly,

148
00:11:17,640 --> 00:11:19,240
not slightly better.

149
00:11:19,240 --> 00:11:23,660
The other thing that's interesting is the paper they released with all of their performance

150
00:11:23,660 --> 00:11:24,660
on the benchmarks.

151
00:11:24,660 --> 00:11:25,660
I do love the...

152
00:11:25,660 --> 00:11:29,160
There's been a few things floating around on the Twitters of, wouldn't it be nice if

153
00:11:29,160 --> 00:11:33,560
we could all mark our own homework in that way?

154
00:11:33,560 --> 00:11:38,520
Obviously, these are standard benchmarking processes, but there are a few little nuances

155
00:11:38,520 --> 00:11:42,800
that people have untangled in there in terms of the prompting techniques and the number

156
00:11:42,800 --> 00:11:48,720
of prompts used to get a better result for Gemini over GPT-4 that are, again, just a little

157
00:11:48,720 --> 00:11:49,720
bit questionable.

158
00:11:49,720 --> 00:11:54,000
So I'm looking at this and thinking it's probably GPT-4 level, but what I really want to know

159
00:11:54,000 --> 00:12:01,480
is what applications are really going to benefit from a model that was trained multimodal first.

160
00:12:01,480 --> 00:12:06,320
That could be a bit of magic in there because it's not clear to me because now I don't trust

161
00:12:06,320 --> 00:12:11,800
any of the demos, how that will translate into real world applications that GPT-4 with

162
00:12:11,800 --> 00:12:17,520
all of its add-ons like Vision and Dali can't do.

163
00:12:17,520 --> 00:12:18,520
I want to see that.

164
00:12:18,520 --> 00:12:19,520
I don't know.

165
00:12:19,520 --> 00:12:20,520
It'll be fun.

166
00:12:20,520 --> 00:12:26,880
There are other really impressive things in the demo as well in the chat interface.

167
00:12:26,880 --> 00:12:33,560
So they showed, it was a chat-based, you know, GPT, chat-GPT style interface, a little barred,

168
00:12:33,560 --> 00:12:44,400
call it whatever you want, where the interface changed because the model itself created new

169
00:12:44,400 --> 00:12:48,440
UI elements in the chat.

170
00:12:48,440 --> 00:12:57,840
So the example it gave was somebody uploading homework with four maths questions on it.

171
00:12:57,840 --> 00:12:59,480
And they've handwritten the responses.

172
00:12:59,480 --> 00:13:04,560
They've taken a photo of the paper, uploaded it, four questions in these blocks.

173
00:13:04,560 --> 00:13:11,160
And then the chat says, which one of these do you want to discuss?

174
00:13:11,160 --> 00:13:12,640
Click on the option.

175
00:13:12,640 --> 00:13:17,640
And it's taken the photo and put on interactive clickable elements on top of it.

176
00:13:17,640 --> 00:13:22,320
And then that allows you to interact with it in that way.

177
00:13:22,320 --> 00:13:28,040
It also showed a demo where very similar to the copilot, I don't know what it is about

178
00:13:28,040 --> 00:13:38,480
the big tech companies, that they love using birthday party planning as a default use case

179
00:13:38,480 --> 00:13:40,520
for using chat LLMs.

180
00:13:40,520 --> 00:13:44,920
But they talked about planning a birthday party for someone and trying to come up with

181
00:13:44,920 --> 00:13:47,680
a theme and ideas.

182
00:13:47,680 --> 00:13:55,600
And the chat does that, comes up with these ideas, but then brings up, it's almost like

183
00:13:55,600 --> 00:14:00,240
in the, like when you do a Google search, it's got the image across the top or the box

184
00:14:00,240 --> 00:14:02,720
of images across the top and you can scroll through it.

185
00:14:02,720 --> 00:14:05,140
And then you've got the knowledge graph at the side.

186
00:14:05,140 --> 00:14:10,400
So it kind of designs in real time, these different UI elements that you can interact

187
00:14:10,400 --> 00:14:12,200
with to say, oh, I like that idea.

188
00:14:12,200 --> 00:14:15,600
You've told me I could have a dinosaur theme for my kid's birthday party.

189
00:14:15,600 --> 00:14:20,120
Click on that and then it expands the chat in a new way and brings up more elements.

190
00:14:20,120 --> 00:14:24,560
So that was really interesting because that's all supposedly done in real time on the fly

191
00:14:24,560 --> 00:14:26,960
by the model itself.

192
00:14:26,960 --> 00:14:32,600
That is interesting because I imagine more of worlds where we speak to computers like

193
00:14:32,600 --> 00:14:35,960
you and I are speaking now and then they do stuff for us.

194
00:14:35,960 --> 00:14:38,920
Mostly because I've been conditioned by Star Trek to think that that's what the future

195
00:14:38,920 --> 00:14:39,920
looks like probably.

196
00:14:39,920 --> 00:14:44,080
But also there's an element of that's how humans interact.

197
00:14:44,080 --> 00:14:49,880
So would that be the best way to interact with a synthetic human or an AI computer?

198
00:14:49,880 --> 00:14:55,360
But I wonder if we're seeing some hints there as to where they think their web-based experience

199
00:14:55,360 --> 00:15:02,400
that we currently call search might evolve into this sort of dynamic interactive tool

200
00:15:02,400 --> 00:15:07,720
that over time they try and find clever ways to put ads in amongst it.

201
00:15:07,720 --> 00:15:12,200
Because you know those generative overlays that you were talking about and in some areas

202
00:15:12,200 --> 00:15:17,840
I think one of them was like a person's making a cake and they're like I want to see cat-based

203
00:15:17,840 --> 00:15:19,160
cakes and a bunch of them come in.

204
00:15:19,160 --> 00:15:21,240
You can scroll through them.

205
00:15:21,240 --> 00:15:26,960
What if every fourth one of those is some sort of sponsored content or an ad to buy

206
00:15:26,960 --> 00:15:30,080
the necessary materials to make said cake or I don't know right.

207
00:15:30,080 --> 00:15:33,160
So yeah I think the jury's out.

208
00:15:33,160 --> 00:15:38,040
I like to be, last couple of weeks I played the role of naysayer a little bit on the podcast

209
00:15:38,040 --> 00:15:45,760
but the truth is we need a model in town that's as good as GPT-4 to keep innovation flying

210
00:15:45,760 --> 00:15:47,440
out the door at open AI.

211
00:15:47,440 --> 00:15:50,640
If you're a user, if you're a tech optimist right because they need that pressure on them

212
00:15:50,640 --> 00:15:52,240
and I think now they have it.

213
00:15:52,240 --> 00:15:56,400
The reason I like to bash Google a bit is they've had and continue to have some of the

214
00:15:56,400 --> 00:16:01,520
brightest AI minds in the world and not like they kick-started this transformer you know

215
00:16:01,520 --> 00:16:06,720
generative pre-transformer revolution and I just I want them to be in the game because

216
00:16:06,720 --> 00:16:12,420
I think they create an ecosystem that moves at speed and that good or real that's kind

217
00:16:12,420 --> 00:16:13,420
of what I want.

218
00:16:13,420 --> 00:16:18,880
I want the ecosystem to move at speed provided you know the terminators don't descend on

219
00:16:18,880 --> 00:16:22,360
us and or turn the whole universe into paper clips.

220
00:16:22,360 --> 00:16:27,840
Well they're very clearly putting the heavy hitters on this because obviously Demis Asabis

221
00:16:27,840 --> 00:16:35,160
is working on this Sundar Pichai was very much involved but also co-founder Sergey Brin

222
00:16:35,160 --> 00:16:42,040
was actually named as a contributor on the paper that they published and I heard someone

223
00:16:42,040 --> 00:16:49,360
talking about they were at an event recently where he was he was there and Brin was speaking

224
00:16:49,360 --> 00:16:56,720
to some AI developers talking to them about some of the decisions that they made with

225
00:16:56,720 --> 00:17:00,840
the development of this particular model but getting really really granular in the detail

226
00:17:00,840 --> 00:17:07,840
like a level that you'd have to be really fully immersed and all over the detail to

227
00:17:07,840 --> 00:17:10,520
even know to ask that question.

228
00:17:10,520 --> 00:17:17,320
So I think that he's been actually coding some of the model.

229
00:17:17,320 --> 00:17:19,320
Nice!

230
00:17:19,320 --> 00:17:25,160
Well I think the summary here for marketers is there's a new game in town and there's

231
00:17:25,160 --> 00:17:28,720
going to be two parts to this game. The first part is we're going to have access to a GPT-4

232
00:17:28,720 --> 00:17:33,080
like model which has got some super skills. It's not super clear yet how we're going to

233
00:17:33,080 --> 00:17:37,920
leverage that but if you think about Google Duet and these other tools that can access

234
00:17:37,920 --> 00:17:43,760
your emails and your files one of the major criticisms we had of those tools including

235
00:17:43,760 --> 00:17:48,760
Bard is that the model that underpin them just wasn't good enough to do anything interesting.

236
00:17:48,760 --> 00:17:54,520
It got too many stuff wrong it was hallucinating all the time whereas you can imagine if you're

237
00:17:54,520 --> 00:17:59,520
a Google workspace company and your Bard and your Google Docs integration and your Google

238
00:17:59,520 --> 00:18:05,160
Sheets and your email integration is powered by Gemini Ultra now we're looking at some

239
00:18:05,160 --> 00:18:10,560
superpowers right so that could be really interesting and a true competitor to Copilot.

240
00:18:10,560 --> 00:18:15,560
So 2024 is going to be interesting for businesses and how they leverage generative AI across

241
00:18:15,560 --> 00:18:20,280
their software suites especially regardless of if they're a Microsoft customer or a Google

242
00:18:20,280 --> 00:18:24,400
customer they're going to have an AI model about the same level that's what it looks

243
00:18:24,400 --> 00:18:34,480
like isn't it. Cool let's talk prompt engineering. So recently Microsoft has been exploring prompt

244
00:18:34,480 --> 00:18:41,640
engineering and revealed some significant advancements in the capabilities of the LLMs

245
00:18:41,640 --> 00:18:47,280
that we know and love like GPT-4. At the heart of this discovery was Medprompt which was

246
00:18:47,280 --> 00:18:52,200
a strategy that significantly enhanced the performance of these models for specific tasks.

247
00:18:52,200 --> 00:18:57,880
So the Medprompt it was initially developed for medical challenges as they probably suggest

248
00:18:57,880 --> 00:19:04,360
but it's proven its versatility by achieving breaking record breaking scores on the MMLU

249
00:19:04,360 --> 00:19:10,040
benchmark which is a test spanning diverse subjects from maths to history. So this is

250
00:19:10,040 --> 00:19:16,520
a series of medical challenge type prompts that are now doing well in other areas by

251
00:19:16,520 --> 00:19:21,320
fine-tuning the prompts given to GPT-4 researchers could steer the AI to understand and respond

252
00:19:21,320 --> 00:19:27,080
with greater accuracy and depth. This method involves combining simple and complex queries

253
00:19:27,080 --> 00:19:33,920
allowing the model to integrate multiple responses and choose the most confident answer. I know

254
00:19:33,920 --> 00:19:39,720
you saw this this week Martin what was most interesting to you about this story? I think

255
00:19:39,720 --> 00:19:43,900
some of the developments around also the techniques that they use in the prompt engineering to

256
00:19:43,900 --> 00:19:51,320
get the outputs here are really interesting. They've effectively described three different

257
00:19:51,320 --> 00:20:01,880
prompting techniques that you can use in order to improve the reliability of the outputs.

258
00:20:01,880 --> 00:20:07,880
So some of them we already know about one of them is I'm going to keep these fairly

259
00:20:07,880 --> 00:20:12,640
simple because they're slightly more technical but one of them is chain of thought prompting

260
00:20:12,640 --> 00:20:23,920
which is the technique of asking the model to explain his reasoning before it kind of

261
00:20:23,920 --> 00:20:28,360
gives you the answer. So talk through its steps equivalent to asking a student to show

262
00:20:28,360 --> 00:20:33,560
it's working. That's something that's been around and been known by researchers for a

263
00:20:33,560 --> 00:20:45,320
while. Another one of the techniques was to where we had few shot prompting so the idea

264
00:20:45,320 --> 00:20:52,640
of in your initial input prompt you give it a couple of examples or a few examples of

265
00:20:52,640 --> 00:20:58,760
the sort of response that you're after or what what good looks like. So you might put

266
00:20:58,760 --> 00:21:03,000
here's you might be a question and answer and you give it several examples of questions

267
00:21:03,000 --> 00:21:10,120
and answers in the format and it will stick to that. Well they've taken a step further

268
00:21:10,120 --> 00:21:20,360
and created something called dynamic few shot prompting which is slightly more technical

269
00:21:20,360 --> 00:21:28,720
and requires the use of an embeddings model in order to be able to identify the best examples

270
00:21:28,720 --> 00:21:36,720
for your few shot prompts but that was a great technique for improving the quality of the

271
00:21:36,720 --> 00:21:47,240
outputs. Yeah and then there was a third technique which this one's this feels like a slight

272
00:21:47,240 --> 00:21:55,640
cheat right but it's basically where you repeat the prompt several times and then you get

273
00:21:55,640 --> 00:22:02,720
the answers from the outputs look across those for the ones where the answer is the most

274
00:22:02,720 --> 00:22:09,160
consistent and that's that's your kind of winning response as it were. So it's not like

275
00:22:09,160 --> 00:22:13,220
you're just asking it for one output you're you're basically running the the prompt several

276
00:22:13,220 --> 00:22:19,920
times and where you notice that there is a consistency in the output that is the output

277
00:22:19,920 --> 00:22:26,800
that you should be going with. I did read a research paper that I am not qualified to

278
00:22:26,800 --> 00:22:31,280
understand it will probably do a terrible job of explaining but one of the things that

279
00:22:31,280 --> 00:22:42,880
I learned was when a large language model has the ability to theorize what is the likely

280
00:22:42,880 --> 00:22:48,960
best answer so in other words if behind the scenes it can come up with five or ten answers

281
00:22:48,960 --> 00:22:54,480
to your query and it has some model or mechanism for ranking the quality of those answers and

282
00:22:54,480 --> 00:22:58,400
then choosing the one that's the most appropriate the quality of the answers that you get is

283
00:22:58,400 --> 00:23:04,320
going to approve in general so it sounds to me like a way of achieving that but without

284
00:23:04,320 --> 00:23:07,760
actually having to have the model do it in the background because you're basically forcing

285
00:23:07,760 --> 00:23:13,280
it to do it and so definitely there was something in that paper about outlining the logic about

286
00:23:13,280 --> 00:23:19,280
why that would work so well. I think it's interesting this story as well because they

287
00:23:19,280 --> 00:23:25,680
used this med prompt strategy with GPT-4 didn't they and its score on the MM-LU went up to

288
00:23:25,680 --> 00:23:32,460
like 90 plus percent which was better than Google's Gemini Ultra having previously had

289
00:23:32,460 --> 00:23:39,680
it reported that Ultra had beaten GPT-4 and it does make me reflect when I look at a lot

290
00:23:39,680 --> 00:23:48,160
of the scores and a lot of the benchmarks it's like GPT-4 79.6% Gemini Ultra 80.1 and

291
00:23:48,160 --> 00:23:54,120
it's like how statistically relevant is that difference and when we see stories like this

292
00:23:54,120 --> 00:24:01,600
where the quality of the prompting increases the score you get on the test more potentially

293
00:24:01,600 --> 00:24:10,000
than the model are we in a law of diminishing returns in terms of once we're at 88% and

294
00:24:10,000 --> 00:24:15,720
we get to 90% and then we get to 91% how much better and how much more usable is that? I

295
00:24:15,720 --> 00:24:19,960
don't know because I'm not an expert in this but I think it's a good question particularly

296
00:24:19,960 --> 00:24:26,360
at the front end for the user because these techniques that it's talking about here you

297
00:24:26,360 --> 00:24:31,120
can't really do through the likes of ChatGPT you can do them through the API in clever

298
00:24:31,120 --> 00:24:37,280
stringing together there's some engineering and pipeline workflow elements and kind of

299
00:24:37,280 --> 00:24:42,640
sequencing of the prompting that you need to do in order to get these outputs I don't

300
00:24:42,640 --> 00:24:50,400
think you can just run it in ChatGPT and get this out so for most users this isn't really

301
00:24:50,400 --> 00:24:59,280
super relevant yeah it's I think the thing I take away from this is GPT-4 was good Gemini

302
00:24:59,280 --> 00:25:04,720
Ultra is probably good but until they figure out how to better understand user intent the

303
00:25:04,720 --> 00:25:08,600
quality of the prompt is still going to be quite an important part in terms of getting

304
00:25:08,600 --> 00:25:13,040
a good output and I think there will continue to be clever prompting techniques that can

305
00:25:13,040 --> 00:25:18,520
enhance your ability to get a good result some of which are completely unexpected and

306
00:25:18,520 --> 00:25:26,160
nobody knows why they work like our next story actually Martin on how to make Claw 2.1 a

307
00:25:26,160 --> 00:25:33,360
bit better not a bit better actually a lot better a significant amount better again simple

308
00:25:33,360 --> 00:25:39,680
prompting technique so this was a paper that was published by Anthropic following the announcement

309
00:25:39,680 --> 00:25:49,680
of Claw 2.1 and that massive 200k context window which is about 500 pages of text so

310
00:25:49,680 --> 00:25:58,280
with all of that context available to you you want to be able to have reliable recall

311
00:25:58,280 --> 00:26:04,960
of the information from within that context and that's something that has proven to be

312
00:26:04,960 --> 00:26:12,120
quite inconsistent Gregory Cameron who was a large language model enthusiast has created

313
00:26:12,120 --> 00:26:19,440
this test it's a needle in a haystack test where he inserts a random sentence somewhere

314
00:26:19,440 --> 00:26:26,960
into the middle of a long document and then asks the model to give him an answer that

315
00:26:26,960 --> 00:26:39,120
relates to that that random sentence the team at Anthropic used that test to try to see

316
00:26:39,120 --> 00:26:46,840
the recall capabilities of this 200k context window with Claw 2.1 and they found and this

317
00:26:46,840 --> 00:26:54,400
is just bonkers right so in the in the base test so with no clever engineering just trying

318
00:26:54,400 --> 00:27:02,120
to run this test the model's performance was 27 percent accurate so really not very good

319
00:27:02,120 --> 00:27:09,680
at all like barely better than one in four times it was getting it right however if they

320
00:27:09,680 --> 00:27:17,880
did a simple prompt adjustment they basically added to the start of the Claude assistants

321
00:27:17,880 --> 00:27:26,560
response so they made the Claude assistants response start with here is the most relevant

322
00:27:26,560 --> 00:27:36,600
sentence in the context colon and then got it to retrieve the answer performance improved

323
00:27:36,600 --> 00:27:47,600
to 98 percent accuracy it's an astonishing turnaround so even at the basically performance

324
00:27:47,600 --> 00:27:52,000
drops off as you put that sentence that the kind of mystery sentence further into the

325
00:27:52,000 --> 00:27:57,120
document and you put it into the middle it starts to get really really bad at being able

326
00:27:57,120 --> 00:28:03,640
to effectively recall that but just putting that at the start of the prompt was a absolute

327
00:28:03,640 --> 00:28:10,760
game changer in terms of performance yeah it's it's fascinating because so to like put

328
00:28:10,760 --> 00:28:17,120
this in some context for practical usage one of the awesome things about Claude that's

329
00:28:17,120 --> 00:28:23,640
better than any model in my hand certainly even better than gpt4 is summarizing documents

330
00:28:23,640 --> 00:28:31,960
but that can include company earnings reports it can be long call transcripts for internal

331
00:28:31,960 --> 00:28:37,920
calls or customer research interviews or you know whatever you can think of and if you

332
00:28:37,920 --> 00:28:42,800
can't trust the outputs of those summaries well then you've got to go read the thing

333
00:28:42,800 --> 00:28:48,680
yourself and definitely i've learned i've seen Claude 2.1 be significantly better when

334
00:28:48,680 --> 00:28:53,320
i give it a random call transcript and ask it to summarize the key items from the call

335
00:28:53,320 --> 00:28:58,520
and the actions it does way better but as you said martin it's showing that for very

336
00:28:58,520 --> 00:29:03,720
long bits of information it's good at remembering the beginning and the end but it doesn't it

337
00:29:03,720 --> 00:29:09,040
can't find things in the middle and remember to output those so if a simple one sentence

338
00:29:09,040 --> 00:29:15,800
prompt can have that type of improvement there's practical utility right there in terms of

339
00:29:15,800 --> 00:29:21,920
being able to trust it with those types of applications if anybody wants to play around

340
00:29:21,920 --> 00:29:32,080
with this the Claude chat bot is not the way to go i think try using it in the in the API

341
00:29:32,080 --> 00:29:38,100
console because within the API console you can actually edit the assistance response

342
00:29:38,100 --> 00:29:44,480
so you can put at the start of that here is the most relevant sentence in the context

343
00:29:44,480 --> 00:29:47,960
and make sure that it always starts this response with that and give it all of that context

344
00:29:47,960 --> 00:29:51,880
beforehand yeah talking about the API and the console

345
00:29:51,880 --> 00:29:59,400
Claude have recently released their kind of cool Claude for Google Sheets well i think

346
00:29:59,400 --> 00:30:03,240
haven't they so you can a little bit like some of i think we've spoken about it on the

347
00:30:03,240 --> 00:30:07,360
podcast before there are some Google Chrome plugins floating around that allow you to

348
00:30:07,360 --> 00:30:18,220
basically prompt different models GPT-4 and Claude is included as cells in a spreadsheet

349
00:30:18,220 --> 00:30:25,880
so why is that useful well in back in the day we don't do this anymore but for an example

350
00:30:25,880 --> 00:30:29,680
Martin and I used to copy paste the URLs of interesting stories that we found through

351
00:30:29,680 --> 00:30:37,520
the week into a Google Sheet in column A column B would automatically format a prompt about

352
00:30:37,520 --> 00:30:44,120
those stories and then column C would actually send that prompt to Claude and have Claude

353
00:30:44,120 --> 00:30:48,920
write a summary of the story for us to use as inspiration for creating the script for

354
00:30:48,920 --> 00:30:53,160
the podcast and of course you can then automate and do clever things off the back of that

355
00:30:53,160 --> 00:30:59,200
but now Claude's basically anthropic of made a tool dedicated to doing that with Claude

356
00:30:59,200 --> 00:31:04,360
is that right have I understood that right Martin yeah that's exactly right and they've

357
00:31:04,360 --> 00:31:10,640
given some examples of Claude for Sheets prompting so you can use it for long-form document Q&A

358
00:31:10,640 --> 00:31:15,560
information extraction you can use it for removing personally identifiable information

359
00:31:15,560 --> 00:31:20,720
so if you've got that in a spreadsheet we'll take that out and and edit it out a customer

360
00:31:20,720 --> 00:31:27,040
support chatbot using FAQs an academic tutor prompt chaining they even have function calling

361
00:31:27,040 --> 00:31:32,240
capabilities now as well which they announced as part of Claude 2.1 and there's a whole

362
00:31:32,240 --> 00:31:38,440
resource available online they've got a Claude for Sheets prompting examples workbench where

363
00:31:38,440 --> 00:31:45,080
you can actually go and give it a give it a whirl so if you are looking to get a bit

364
00:31:45,080 --> 00:31:51,960
more gangster in your AI usage and you're trying to think about clever multi-step sequences

365
00:31:51,960 --> 00:31:57,800
or processing lots of information in one go but using the power of AI it's probably worth

366
00:31:57,800 --> 00:32:03,800
going and having a little play and understanding what this new Sheets capability can do right

367
00:32:03,800 --> 00:32:10,600
next story we're going to be talking about Meta's AI image generation tool so Meta's

368
00:32:10,600 --> 00:32:19,440
getting in on the image generation with their imagine with Meta AI tool which utilizes an

369
00:32:19,440 --> 00:32:26,480
image synthesis model trained on over 1.1 billion images from Facebook and Instagram

370
00:32:26,480 --> 00:32:31,280
which is not an insignificant amount of data and of course why we look at a lot of the

371
00:32:31,280 --> 00:32:38,160
companies outside of open AI who have huge treasure troves of data to fuel training their

372
00:32:38,160 --> 00:32:44,560
models on and this is a great example of Meta stroke Facebook stroke Instagram doing exactly

373
00:32:44,560 --> 00:32:49,400
that what the tool will do is it will allow you to generate unique images from written

374
00:32:49,400 --> 00:32:57,080
prompts which was a feature that was previously embedded only in social apps like Instagram

375
00:32:57,080 --> 00:33:03,400
it's kind of interesting in that the models been trained on a vast array of sort of public

376
00:33:03,400 --> 00:33:09,360
social media images and that's a bit of a little bit of an ethical minefield what are

377
00:33:09,360 --> 00:33:10,360
your thoughts on that?

378
00:33:10,360 --> 00:33:15,600
Those terms and conditions sheets that we've all signed up to when we signed up for Instagram

379
00:33:15,600 --> 00:33:21,840
and Facebook all those many years ago and the idea that you know Facebook can use our

380
00:33:21,840 --> 00:33:31,920
data if it wants to well we're now seeing the the fruits of that terms and conditions

381
00:33:31,920 --> 00:33:38,800
authorization I guess yeah so have you used it it's available only in the US but is available

382
00:33:38,800 --> 00:33:41,040
via VPN have you given it a try?

383
00:33:41,040 --> 00:33:44,800
I haven't played with it directly but I've seen a number of side-by-side comparisons

384
00:33:44,800 --> 00:33:50,680
where people feed the same prompt into Dali 3 mid-journey Meta why have you played?

385
00:33:50,680 --> 00:33:57,160
I have and yeah I'm impressed I think it is one of the better models out there certainly

386
00:33:57,160 --> 00:34:03,200
for a first release from a company you know this isn't V3 this is their kind of opening

387
00:34:03,200 --> 00:34:09,840
gambit on the market they've done a really solid job I prompted it with the northern

388
00:34:09,840 --> 00:34:16,440
lights in the skies above the city of Derby and I chose the city of Derby because no other

389
00:34:16,440 --> 00:34:25,080
model can get you know there isn't loads of photos of the AI blast skyline right yeah

390
00:34:25,080 --> 00:34:33,480
you know like you know parochial provincial towns right they're not well known for their

391
00:34:33,480 --> 00:34:38,120
vast quantity of training data in these things and actually did a really good job to the

392
00:34:38,120 --> 00:34:43,960
point where when it generated the images all it generates four at a time and all four of

393
00:34:43,960 --> 00:34:50,040
them were very very similar but I immediately went oh that nearly tricked me into thinking

394
00:34:50,040 --> 00:34:55,720
that skyline was was Derby but I've used it for other things you know the kind of classic

395
00:34:55,720 --> 00:35:01,960
panda riding a bicycle kind of thing and yeah it does a really good job I think they've

396
00:35:01,960 --> 00:35:07,260
absolutely nailed it I think it's supposed to be particularly good for you know selfie

397
00:35:07,260 --> 00:35:12,880
style images as well which wouldn't be a massive surprise given that it was trained on Instagram

398
00:35:12,880 --> 00:35:19,280
right is it what's this photo realism like is there any good yeah I haven't gone I haven't

399
00:35:19,280 --> 00:35:25,080
gone deep into the the different styles um I think there's a way that there is there

400
00:35:25,080 --> 00:35:32,480
is I haven't seen it be truly photorealistic um which I think almost has gone from Dali

401
00:35:32,480 --> 00:35:43,000
3 that capability now there's there's very much a style I think with with Dali 3 it has

402
00:35:43,000 --> 00:35:47,880
its own aesthetic which has been impressed upon it it almost seems like no matter what

403
00:35:47,880 --> 00:35:53,800
I ask it for I'm getting a very similar look and feel but with this model I think they've

404
00:35:53,800 --> 00:36:02,240
done a really good job I can't get photorealism out of Dali 3 for love nor money now it's

405
00:36:02,240 --> 00:36:12,000
um I wonder if a little bit of it's the deep fake aspect right like it almost avoiding

406
00:36:12,000 --> 00:36:17,040
being photorealistic so it's just five percent off looking real so that anybody who looks

407
00:36:17,040 --> 00:36:21,800
at it is going to go well that's AI generated yeah I can only I can only really get good

408
00:36:21,800 --> 00:36:26,000
stuff out of mid-journey still yeah and for photorealism I still think that the images

409
00:36:26,000 --> 00:36:31,520
that you get from Dali are very good I've created some fantastic images on there recently

410
00:36:31,520 --> 00:36:36,320
that I've used in various applications um but if you want photorealism there's that

411
00:36:36,320 --> 00:36:41,420
not even uncanny valley it's just there is a there is something that looks like it's

412
00:36:41,420 --> 00:36:47,280
computer generated it's something that's slightly softer yeah I think it's softer and there's

413
00:36:47,280 --> 00:36:53,040
a there's a feeling of lack of imperfections I can't really describe it like the people's

414
00:36:53,040 --> 00:37:00,400
faces don't look like photos they look like what I would imagine a computer game in 2050

415
00:37:00,400 --> 00:37:05,240
looks like right like the computer generated images on it are so good as to almost look

416
00:37:05,240 --> 00:37:10,800
real but without the imperfections that real humans have not on this podcast obviously

417
00:37:10,800 --> 00:37:18,320
but the other humans yeah the rest of the 8 billion out there slim oh crumbs funny that

418
00:37:18,320 --> 00:37:26,200
would trick um there's some other cool stuff on the AI image meta Instagram Facebook Mary

419
00:37:26,200 --> 00:37:33,800
yeah and um what we're seeing here is that they're starting to incorporate some of the

420
00:37:33,800 --> 00:37:40,520
models that we've spoken about in previous episodes and they're launching these into

421
00:37:40,520 --> 00:37:46,080
consumer applications so they've just announced a new feature on Instagram called backdrop

422
00:37:46,080 --> 00:37:51,280
it's available only in the US at the moment again those of us in the UK and Europe will

423
00:37:51,280 --> 00:38:01,920
just have to rest our souls and patience this model takes the segment anything model which

424
00:38:01,920 --> 00:38:07,440
we talked about a while ago this is basically a computer vision model which can basically

425
00:38:07,440 --> 00:38:15,000
it does like clip paths around object identification but it can do that at a really impressive

426
00:38:15,000 --> 00:38:20,840
scale and it uses that to remove the background and then you can put in any background that

427
00:38:20,840 --> 00:38:26,640
you want so it was similar to what we've seen on things like clip drop and you know back

428
00:38:26,640 --> 00:38:31,320
any kind of AI tool that uses background remover but they've incorporated the segment anything

429
00:38:31,320 --> 00:38:41,640
model and a model called emu so yeah this is the actual rollout and commercialization

430
00:38:41,640 --> 00:38:47,800
of some of their research tools into the Instagram platform yeah I quite like that because we

431
00:38:47,800 --> 00:38:52,640
did talk about that segment anything model I don't know six months ago maybe I think

432
00:38:52,640 --> 00:38:58,960
yeah yeah so what that would imply is that it could take six months to see some of these

433
00:38:58,960 --> 00:39:04,080
research papers make it into products which is not that long really when you think about

434
00:39:04,080 --> 00:39:11,280
it as an ex-scientist I get particularly excited about the segment and anything although whether

435
00:39:11,280 --> 00:39:18,960
it can really segment stuff in biological images like images taken of cells or you know

436
00:39:18,960 --> 00:39:27,520
biopsy samples is another question but a major part of biological research and pathology when

437
00:39:27,520 --> 00:39:33,520
you're doing analysis of a you know like a cancer biopsy sample is image analysis and

438
00:39:33,520 --> 00:39:38,560
segmentation where you have to highlight the interesting part of the image which I you

439
00:39:38,560 --> 00:39:42,880
know there have been AI tools that have emerged and they are really rather good but when I

440
00:39:42,880 --> 00:39:49,480
did my PhD many years ago you had to pretty much do that manually which took hours I think

441
00:39:49,480 --> 00:39:56,640
I've definitely spoken on LinkedIn about taking pictures of lots of fruit fly eggs with fused

442
00:39:56,640 --> 00:40:00,640
dorsal appendages and having to count them by hand and spend a good part of about 10

443
00:40:00,640 --> 00:40:05,000
to 14 days 10 hours a day at the microscope so you could probably automate in about two

444
00:40:05,000 --> 00:40:10,720
hours but it's kind of cool because we talk a lot about marketing and sales and I know

445
00:40:10,720 --> 00:40:13,480
most of the people who are listening they love marketing sales and that's why they're

446
00:40:13,480 --> 00:40:18,240
listening to the podcast but well I do get excited about some of the image analysis capabilities

447
00:40:18,240 --> 00:40:20,400
that some of these tools are going to enable.

448
00:40:20,400 --> 00:40:29,520
Next story is about looking at which is more is capable of running more persuasive sort

449
00:40:29,520 --> 00:40:36,920
of advertising campaigns humans or chat GPT so in a recent study by MIT where they partnered

450
00:40:36,920 --> 00:40:41,800
with a top consulting firm they were trying to figure out who could create the most persuasive

451
00:40:41,800 --> 00:40:48,940
content and there was also a test where it was like AI only humans only or AI and human

452
00:40:48,940 --> 00:40:53,680
working together and the research focused on creative content for five retail products

453
00:40:53,680 --> 00:41:00,560
and five campaign topics using both professional content creators and churchypt. Over 1200

454
00:41:00,560 --> 00:41:04,080
online participants evaluated the content and they were assessing it for things like

455
00:41:04,080 --> 00:41:09,440
satisfaction willingness to pay how interested in it they were how persuasive they found

456
00:41:09,440 --> 00:41:14,440
it and they were categorized into groups with varying levels of awareness about the content's

457
00:41:14,440 --> 00:41:19,800
origin so whether or not they knew it come from a human or an AI or from the two working

458
00:41:19,800 --> 00:41:25,760
together and the key findings were that when content was either created solely or ultimately

459
00:41:25,760 --> 00:41:32,080
determined by chat GPT it actually scored higher in quality across satisfaction willingness

460
00:41:32,080 --> 00:41:38,800
to pay and persuasion metrics compared to purely human generated content so that's interesting

461
00:41:38,800 --> 00:41:45,520
finding number one. When participants knew the content was made only by humans they perceived

462
00:41:45,520 --> 00:41:53,080
it as being of higher quality very what's Elon Musk statement very very no not Elon

463
00:41:53,080 --> 00:41:57,080
Musk Elon Musk was the species wasn't he was very pro-human I'm pro-human as well but it

464
00:41:57,080 --> 00:42:01,880
shows the rest of us are quite pro-human when we know humans have made it we think it has

465
00:42:01,880 --> 00:42:05,920
higher quality although it's fair it's worth highlighting that when people knew it was

466
00:42:05,920 --> 00:42:11,720
AI it didn't diminish the quality of of what people felt they saw it was just if they knew

467
00:42:11,720 --> 00:42:19,760
it was humans they thought it was better the quality gap itself was actually narrower between

468
00:42:19,760 --> 00:42:26,720
AI and humans for product advertising than and campaign messages which I think probably

469
00:42:26,720 --> 00:42:31,480
makes sense and despite a general favoritism towards human generated content for both products

470
00:42:31,480 --> 00:42:39,960
and campaigns there was no consistent aversion to AI included bits and pieces so that's pretty

471
00:42:39,960 --> 00:42:48,600
interesting I think it fits in with this whole is creativity the last bastion of human endeavor

472
00:42:48,600 --> 00:42:54,840
and this is another potential piece of data that says working together we do the best

473
00:42:54,840 --> 00:43:03,040
stuff what do you think Martin? Well it certainly goes to show that AI naysayers the people

474
00:43:03,040 --> 00:43:09,480
who say oh you can't write good copy shows they're wrong but I think we've already seen

475
00:43:09,480 --> 00:43:14,480
this with other research studies that came out recently about the creativity of GPT-4

476
00:43:14,480 --> 00:43:22,200
that is actually it's as creative or more creative than most people and I think that's

477
00:43:22,200 --> 00:43:26,720
it's probably not as good as the in like in this example for instance I would love to

478
00:43:26,720 --> 00:43:35,460
see the the best copywriters in the world pitched against GPT-4 I would well imagine

479
00:43:35,460 --> 00:43:45,680
that the best copywriters are better still but against the average kind of copywriting

480
00:43:45,680 --> 00:43:55,680
team or marketing team GPT-4 can perform as well or better than those but again doing

481
00:43:55,680 --> 00:44:03,160
it in conjunction with humans is where we see strong results so I think there's nobody's

482
00:44:03,160 --> 00:44:08,120
going to be completely out of a job just yet in fact there was a something I read this

483
00:44:08,120 --> 00:44:16,160
week where some big industry AI heads were were saying that 80% of 80% of jobs will be

484
00:44:16,160 --> 00:44:23,040
impacted by AI but no jobs will be replaced entirely and it's just going to change the

485
00:44:23,040 --> 00:44:29,840
way that we all work and it augment the way that we work right? Yeah that definitely fits

486
00:44:29,840 --> 00:44:34,920
with my sense I mean when you and I are out and about speaking at conferences we get an

487
00:44:34,920 --> 00:44:41,540
asked a version of this question pretty much all the time yeah and lo and behold I was

488
00:44:41,540 --> 00:44:48,120
asked several times when I was at SAHPS last week and I think what you touched on is very

489
00:44:48,120 --> 00:44:53,200
much in line with a model that Ethan Molek talks a lot about we talk about Ethan a lot

490
00:44:53,200 --> 00:44:57,160
on the podcast I think it was related to the paper they published where Boston Consulting

491
00:44:57,160 --> 00:45:03,400
Group had two groups of consultants one used GPT-4 and one didn't and in general the team

492
00:45:03,400 --> 00:45:07,800
that used GPT-4 on this fictional project they were working on were faster more efficient

493
00:45:07,800 --> 00:45:12,880
and there was a higher quality of their outputs but when you dive a bit into the data it's

494
00:45:12,880 --> 00:45:18,400
almost like it squeezed the normal distribution curve of capability and what I mean by that

495
00:45:18,400 --> 00:45:23,580
is if you're a super expert you probably didn't see your quality increase that much when you

496
00:45:23,580 --> 00:45:28,160
had access to GPT-4 but if you're in the middle of the normal distribution curve or perhaps

497
00:45:28,160 --> 00:45:33,160
let's say even a slightly weaker member of the team it helped accelerate and improve

498
00:45:33,160 --> 00:45:40,920
the quality of your output significantly so I do get the sense to your point about how

499
00:45:40,920 --> 00:45:45,040
would the master copywriters get on against GPT-4 I think they would absolutely smash

500
00:45:45,040 --> 00:45:49,760
it out of the war to be honest but I think you're right I think it can help people who

501
00:45:49,760 --> 00:45:56,040
don't have 30 years of experience as ninja copywriters you know 10,000 hours or whatever

502
00:45:56,040 --> 00:46:01,360
it is they say you've got to have of just refining your craft maybe with only 500 hours

503
00:46:01,360 --> 00:46:07,000
you can do the output of someone who's had 8,000 hours worth of training which to a certain

504
00:46:07,000 --> 00:46:11,760
extent certainly in the short to medium term is going to be really really valuable I think

505
00:46:11,760 --> 00:46:18,640
if you model that out further I think the super duper uber experts become even more

506
00:46:18,640 --> 00:46:22,640
valuable because now we've got this nice fat middle where everything's much better than

507
00:46:22,640 --> 00:46:28,860
it used to be but then now the bar the average is higher and so to stand out above that average

508
00:46:28,860 --> 00:46:34,260
you need the ninjas and I think it's those super duper copywriters that will continue

509
00:46:34,260 --> 00:46:38,080
to get paid the big bucks because they'll be able to produce the quality of work that

510
00:46:38,080 --> 00:46:44,340
even all the rest of us with GPT-4 in hand just won't be able to do so if you're listening

511
00:46:44,340 --> 00:46:50,240
to this and you're thinking what should I do in the realms of the AI world firstly get

512
00:46:50,240 --> 00:46:54,760
to grips with the tools learn about them test them figure out how they can help you be better

513
00:46:54,760 --> 00:47:00,000
at what you do and then continue to invest in developing your domain level expertise

514
00:47:00,000 --> 00:47:06,320
to an absolute uber area where it will take well hopefully a long time for any AI model

515
00:47:06,320 --> 00:47:09,960
to be really really good at the thing that you've spent ages getting good at because

516
00:47:09,960 --> 00:47:16,480
those models are still somewhat generalists but also don't be surprised if the model catches

517
00:47:16,480 --> 00:47:22,120
up and completely crushes your dreams within six months because that's the brutal reality

518
00:47:22,120 --> 00:47:26,520
that we all now live in so I don't know if you saw but I was looking at you Martin to

519
00:47:26,520 --> 00:47:33,920
say don't say that Martin don't say it yeah okay I do think it could be true but it is

520
00:47:33,920 --> 00:47:38,900
Christmas time and we're trying back in my optimist optimistic approach because it could

521
00:47:38,900 --> 00:47:44,300
be three months yeah it could be our Gemini Ultra came out crumbs we saw a rumor this

522
00:47:44,300 --> 00:47:48,040
week we were going to talk about this but I think we should there was a rumor about

523
00:47:48,040 --> 00:47:53,760
GPT 4.5 floating around the Twittersphere wasn't there Martin and all the different

524
00:47:53,760 --> 00:47:59,880
things that it could do video 3d the 3d I didn't even say 3d what one assumed some sort

525
00:47:59,880 --> 00:48:07,120
of 3d imagery or can it create maistime continuums like who knows what's going on the cost per

526
00:48:07,120 --> 00:48:12,120
token was astronomical I don't know how anyone could run a business off the back of it and

527
00:48:12,120 --> 00:48:17,400
but we think it's fake don't we mine but if GPT 4.5 came out and was like even better

528
00:48:17,400 --> 00:48:24,240
than Gemini where would we be and when I was like when you send that to me and I was looking

529
00:48:24,240 --> 00:48:31,400
at I was trying to think there's a big challenge at the moment about defining AGI right advanced

530
00:48:31,400 --> 00:48:35,680
general intelligence a computer that can do most things better than a human and I read

531
00:48:35,680 --> 00:48:42,400
a really interesting take on this that if GPT 4 if chat GPT as we will use it today

532
00:48:42,400 --> 00:48:50,040
came out 10 years ago we'd be calling it AGI which I thought was an interesting and provocative

533
00:48:50,040 --> 00:48:55,460
point right it's like the more sophisticated our tools get the higher we keep raising the

534
00:48:55,460 --> 00:49:02,880
bar of AGI to a bar where I feel like the new bar is nobody has to work ever again that's

535
00:49:02,880 --> 00:49:13,840
why I feel like the AGI bar is well which actually chimes somewhat with Sam Altman's

536
00:49:13,840 --> 00:49:21,080
interview with Time magazine this week and actually in that he talks about the next the

537
00:49:21,080 --> 00:49:28,720
next steps for AI in terms of productivity and well people having more time and more

538
00:49:28,720 --> 00:49:34,440
freedom to be able to do what they want because he's you can see that they're clearly thinking

539
00:49:34,440 --> 00:49:44,200
at OpenAI where this is going is it is going to well agents when AI can really effectively

540
00:49:44,200 --> 00:49:51,000
execute tasks and everybody goes from having chat GPT which in this interview he describes

541
00:49:51,000 --> 00:49:54,480
as I can't remember his exact phrase but it was something along the lines of it's a bit

542
00:49:54,480 --> 00:50:04,840
rubbish to and he's like you know fast forward a year or two's time everybody's got 10 AIs

543
00:50:04,840 --> 00:50:11,960
doing tasks for them doing work for them in their pocket 24 hours a day 7 days a week

544
00:50:11,960 --> 00:50:20,600
to 5 years time when everyone's got hundreds of AIs doing hundreds of tasks for them 24

545
00:50:20,600 --> 00:50:24,600
hours a day 7 days a week and what does that allow you to do from a personal productivity

546
00:50:24,600 --> 00:50:28,600
perspective and what does that do to the economy?

547
00:50:28,600 --> 00:50:35,040
Yeah that is I think that's the thing that keeps I think that's probably what keeps people

548
00:50:35,040 --> 00:50:42,280
up who are in these circles more than AI accidentally turns the universe into paper clips because

549
00:50:42,280 --> 00:50:47,080
it's been told to make the best paper clip but wasn't given any common sense briefing

550
00:50:47,080 --> 00:50:51,240
instructions and so just keeps churning them out until the universe is all paper clips

551
00:50:51,240 --> 00:50:59,120
or terminator right I think it's how does it affect the economy when the main force

552
00:50:59,120 --> 00:51:05,680
of the economy is still human labour we're still paid to work what happens when you unshackle

553
00:51:05,680 --> 00:51:08,000
output from human input.

554
00:51:08,000 --> 00:51:14,240
Ilya Sutsgiver in an interview a few months ago talked about economic destruction that

555
00:51:14,240 --> 00:51:21,960
was that was the phrase he used with AGI that's all he's used sometimes though you know like

556
00:51:21,960 --> 00:51:26,320
if there's a if there's a fire in a forest sometimes that's needed because some of the

557
00:51:26,320 --> 00:51:30,960
dead trees you know they're destroyed it gives a chance for the saplings to grow up and the

558
00:51:30,960 --> 00:51:36,380
forest is refreshed and it probably takes some time but ultimately it's a can be a productive

559
00:51:36,380 --> 00:51:41,880
and natural process so I think it's whether or not he means doom and gloom destruction

560
00:51:41,880 --> 00:51:48,040
or creative destruction yeah but it'll be definitely like the ways that it wipes out

561
00:51:48,040 --> 00:51:52,680
the dinosaurs right and allow the conditions for mammals to flourish without which we wouldn't

562
00:51:52,680 --> 00:51:59,040
be here today so maybe that's the exciting future we've all got to look forward to no

563
00:51:59,040 --> 00:52:02,280
more dinosaurs are you a dinosaur Paul?

564
00:52:02,280 --> 00:52:08,080
Well that dinosaur turn of phrase could be used in another way couldn't it in the in

565
00:52:08,080 --> 00:52:14,960
the rise of AI but I'm sure when robot Paul and robot Martin are talking about how sadly

566
00:52:14,960 --> 00:52:21,320
the humans made way for the more evolved efficient effective robotic life forms that they gave

567
00:52:21,320 --> 00:52:27,960
birth to I wonder if they'll if they'll feel bad about the humans like do you ever feel

568
00:52:27,960 --> 00:52:32,320
sorry for the dinosaurs like oh bless them they got wiped out by an asteroid but we did

569
00:52:32,320 --> 00:52:36,760
all right out of it yeah well I mean they'll just turn into chickens

570
00:52:36,760 --> 00:52:46,920
well not all of them but yeah the yeah look ultimately progress is progressing as it is

571
00:52:46,920 --> 00:52:52,600
and without we get too deep into the weeds of where this all might go because I don't

572
00:52:52,600 --> 00:52:57,600
think anybody really knows either that's the other thing all we can do is marketing professionals

573
00:52:57,600 --> 00:53:03,400
and sales professionals is keep doing the best we can to have an impact enjoy our work

574
00:53:03,400 --> 00:53:08,280
if these tools can help with that then you know better to be in the gang figuring out

575
00:53:08,280 --> 00:53:11,440
how to use these tools I think than sticking your head in the sand

576
00:53:11,440 --> 00:53:15,240
should we move on to our last couple of stories?

577
00:53:15,240 --> 00:53:20,240
Yeah one that is very much rooted in the world of people who know a thing or two about marketing

578
00:53:20,240 --> 00:53:29,040
that is McDonald's and Google put our press release basically announcing that McDonald's

579
00:53:29,040 --> 00:53:37,800
is going to be using Google Cloud as its main cloud provider and whilst the press release

580
00:53:37,800 --> 00:53:42,800
in and of itself isn't particularly interesting in the headline it did talk about generative

581
00:53:42,800 --> 00:53:49,920
AI so that's what turned me on to it and in the press release they say we're this new

582
00:53:49,920 --> 00:53:56,000
partnership a dedicated Google Cloud team in Chicago will work in close proximity to

583
00:53:56,000 --> 00:54:03,240
McDonald's global innovation center known as Speedy Labs together they'll focus on applying

584
00:54:03,240 --> 00:54:08,080
generative AI across a number of key business priorities to power exciting new experiences

585
00:54:08,080 --> 00:54:14,140
for crew and customers with McDonald's unmatched convenience and value.

586
00:54:14,140 --> 00:54:21,440
Now we know McDonald's likes data and digital innovation when they introduced their self-service

587
00:54:21,440 --> 00:54:32,960
kiosks they were reported to have increased order size by 30% and order value by 20% and

588
00:54:32,960 --> 00:54:38,960
I heard something the other day that said they found that men when using kiosks about

589
00:54:38,960 --> 00:54:46,520
20% more or something order two burgers as opposed to one because they don't have to

590
00:54:46,520 --> 00:54:50,840
say to someone at the desk or gym because they're thinking that person's thinking I'm

591
00:54:50,840 --> 00:54:57,600
a proper fat they're judging me they're judging me yep and so that increases average spend

592
00:54:57,600 --> 00:55:02,600
so we know that they're they're into their their data and how digital can transform the

593
00:55:02,600 --> 00:55:08,520
customer experience it got me thinking when you've got those kiosks and those digital

594
00:55:08,520 --> 00:55:17,640
displays and people are using the my McDonald's reward to connect their data in real time

595
00:55:17,640 --> 00:55:22,400
to this digital display so the display knows exactly who you are knows everything about

596
00:55:22,400 --> 00:55:31,680
you is hooked up to the cloud the potential for generative AI to come up with personalized

597
00:55:31,680 --> 00:55:36,480
interactions there is really quite remarkable you can imagine it from personalized menu

598
00:55:36,480 --> 00:55:44,160
options which in and of itself isn't particularly exciting to avatars that you know almost like

599
00:55:44,160 --> 00:55:48,840
a hey Jen style avatars that will take your order in real time whether or not people need

600
00:55:48,840 --> 00:55:54,040
that and actually just tapping on the burger that you want and you know hammering the plus

601
00:55:54,040 --> 00:55:59,800
sign till you get to the relevant number of burgers that you desire in your in your takeout

602
00:55:59,800 --> 00:56:03,920
whether that's the way to go about it or not I don't know but yeah just made me think about

603
00:56:03,920 --> 00:56:08,280
the potential here this is obviously a big deal right McDonald's has huge amounts of

604
00:56:08,280 --> 00:56:14,520
customer data on their my rewards at plug that into Google cloud and bring that down

605
00:56:14,520 --> 00:56:19,520
into a digital interface where people are actually spending their cash you can you can

606
00:56:19,520 --> 00:56:26,280
see there could be quite the uplift in average customer order yeah it's interesting the first

607
00:56:26,280 --> 00:56:31,480
thing is thank you for the strategy to go and order more burgers without feeling like

608
00:56:31,480 --> 00:56:36,080
I'm being judged because that was an application I needed in my life and I'm definitely taking

609
00:56:36,080 --> 00:56:42,080
away from this conversation the other thing is I think on the podcast we try to go into

610
00:56:42,080 --> 00:56:46,520
as many first order impacts of things as we can and occasionally we get a bit deeper into

611
00:56:46,520 --> 00:56:52,560
second order impacts right but the smart people are playing these things out right they're

612
00:56:52,560 --> 00:56:58,360
like well if this enables this and this enables this and this enables this how do they all

613
00:56:58,360 --> 00:57:05,500
come together to impact my business right if you are McDonald's how does the ability

614
00:57:05,500 --> 00:57:11,440
to leverage all the data you have on customers combine with synthetic humans combine with

615
00:57:11,440 --> 00:57:16,040
predictive modeling to know exactly what type of thing people are going to order combined

616
00:57:16,040 --> 00:57:24,360
with tests in using generative AI tools that make the outlay of the screen and everything

617
00:57:24,360 --> 00:57:28,600
ever more persuasive to get you to buy more like how do all of those things come together

618
00:57:28,600 --> 00:57:32,320
to influence your business and like I say there'll be people smarter certainly than

619
00:57:32,320 --> 00:57:38,760
me modeling those out and I think they're they will potentially end up being the unexpected

620
00:57:38,760 --> 00:57:44,240
success stories of where AI gets used in a way that we probably as marketers we're thinking

621
00:57:44,240 --> 00:57:49,880
yeah I can write blog posts and social posts and now I can turn Martin's voice in from

622
00:57:49,880 --> 00:57:54,680
his voice into a strict older woman but what how do these things does it call back to a

623
00:57:54,680 --> 00:57:58,280
previous episode as well so if you think what the heck is he talking about that you'll have

624
00:57:58,280 --> 00:58:03,120
to listen to episode 35 I think it is but I think it's going to be the ones that figure

625
00:58:03,120 --> 00:58:06,920
out how to connect that stuff to do unexpected things that drive a lot commercial value and

626
00:58:06,920 --> 00:58:11,200
perhaps that's what McDonald's play is here.

627
00:58:11,200 --> 00:58:16,240
Multivariate testing right something where you can run lots of different examples of

628
00:58:16,240 --> 00:58:20,560
an interface or a scenario or what have you and you see that they do market testing of

629
00:58:20,560 --> 00:58:25,120
different burgers across the different country right to see what's going to land well well

630
00:58:25,120 --> 00:58:32,040
and you can have as we discussed earlier Google Gemini is able to in real time dynamically

631
00:58:32,040 --> 00:58:39,320
update the user interface element that applies to that kiosk where it is doing multivariate

632
00:58:39,320 --> 00:58:47,480
testing on names of products or combos of meals dynamically creating different mix and

633
00:58:47,480 --> 00:58:54,120
match menu items or what have you and it's doing this all autonomously doing you know

634
00:58:54,120 --> 00:59:02,120
A B testing market testing real time on steroids like never seen before that's yeah that's

635
00:59:02,120 --> 00:59:09,760
going to be more impactful on business than me speaking like a strict older woman.

636
00:59:09,760 --> 00:59:14,240
Depends what in what context that is you but I agree I guess yeah I do agree with you and

637
00:59:14,240 --> 00:59:23,920
honestly if I got if I chopped off to McDonald's moving on if I got to McDonald's kiosk if

638
00:59:23,920 --> 00:59:28,960
I had facial recognition and it knew I was and I didn't have to press 18 button presses

639
00:59:28,960 --> 00:59:34,600
to get the meal that I order 95% of the time I know some people would find that a bit like

640
00:59:34,600 --> 00:59:39,240
invasive but personally I'd be like brilliant just saved me eight minutes of pressing the

641
00:59:39,240 --> 00:59:44,840
wrong buttons just pop up synthetic person says hi Paul how you doing do you want yeah

642
00:59:44,840 --> 00:59:50,440
do you want eight burgers three chips and 16 tubs of mayonnaise and I'm like yeah just

643
00:59:50,440 --> 00:59:55,320
like last time the family they're all yeah the extended family they're all they're just

644
00:59:55,320 --> 00:59:59,600
yeah so yeah I'd be up for that but yeah I think you're right that multivariate testing

645
00:59:59,600 --> 01:00:05,160
would be be interesting anyway we better move on so at the hour mark give or take and we've

646
01:00:05,160 --> 01:00:15,600
got one more story it is about an open source model called Mistral so Mistral is a French

647
01:00:15,600 --> 01:00:23,600
company that received another round of funding this week I think they're only about six months

648
01:00:23,600 --> 01:00:28,120
old but the latest round of funding puts their valuation at the two billion dollar mark so

649
01:00:28,120 --> 01:00:33,800
they're another example of money rushing into the AI space into certain areas but why are

650
01:00:33,800 --> 01:00:37,720
Mistral interesting well they're interesting for a number of reasons but this week they

651
01:00:37,720 --> 01:00:45,200
became even more interesting when they released their new model which is mixed raw eight times

652
01:00:45,200 --> 01:00:54,400
seven B which is a mixture of experts model so it's pretty cool because it's open source

653
01:00:54,400 --> 01:00:58,800
which will come on to why that's important in a moment but just as importantly it has

654
01:00:58,800 --> 01:01:08,880
beaten the likes of GPT 3.5 and met as Lama Lama 2 on a number of benchmarks so now you've

655
01:01:08,880 --> 01:01:14,780
got a free open source model that doesn't come with all of the limitations that comes

656
01:01:14,780 --> 01:01:23,200
with all of the reinforcement aligned models all of the weights of the models are all locked

657
01:01:23,200 --> 01:01:29,760
down you can't change them that you get with say GPT 3.5 but you can now do that with mixed

658
01:01:29,760 --> 01:01:36,280
raw eight times seven B it's also interesting that it is a mixture of experts model which

659
01:01:36,280 --> 01:01:42,360
is what people think is the underlying architecture of GPT 4 so rather than just like making a

660
01:01:42,360 --> 01:01:48,320
bigger and bigger and bigger and bigger and bigger model that the mixed raw eight times

661
01:01:48,320 --> 01:01:56,000
seven B has this mixture of experts model it's also got far fewer parameters than GPT

662
01:01:56,000 --> 01:02:04,080
3.5 I think it's about a third isn't it Martin so yeah 46 billion parameters compared to

663
01:02:04,080 --> 01:02:12,000
about 175 billion parameters for GPT 3.5 so you get that level of GPT 3.5 performance

664
01:02:12,000 --> 01:02:18,040
but about 25 30 percent of the model size this is important right because model running

665
01:02:18,040 --> 01:02:22,480
large models is expensive storing them is expensive if you ever want to be able to run

666
01:02:22,480 --> 01:02:27,400
a half decent model on your phone those model sizes are going to need to come down so this

667
01:02:27,400 --> 01:02:33,720
is also an example of what happens when you come up with clever ways to train really good

668
01:02:33,720 --> 01:02:37,760
models but you compress the information even more so the models are smaller and therefore

669
01:02:37,760 --> 01:02:43,720
a bit cheaper to run the I think the critical aspect of this is it's open source so you

670
01:02:43,720 --> 01:02:50,240
can run it on your own machine there are still lots of concerns about what happens when you

671
01:02:50,240 --> 01:02:55,880
share data with open AI Microsoft Google like everybody's happy sharing their data with

672
01:02:55,880 --> 01:02:59,400
their company until something bad happens and they wish they hadn't so for those that

673
01:02:59,400 --> 01:03:05,480
are super paranoid about that the ability to get ever better model performance running

674
01:03:05,480 --> 01:03:10,560
on say your own laptop even if it's somewhat slowly unless you've got a super dupe powerful

675
01:03:10,560 --> 01:03:18,640
PC with cool GPUs in it or a nice powerful Mac it's going to give people access and open

676
01:03:18,640 --> 01:03:25,680
up use cases that you just can't get with the other models so kind of cool.

677
01:03:25,680 --> 01:03:33,080
It is very cool and I'm always interested to see where the open source community is

678
01:03:33,080 --> 01:03:41,360
going they had another bonus recently and so much as the EU's AI Act regulation isn't

679
01:03:41,360 --> 01:03:47,560
going to cover open source models in the same way that it's covering the closed models so

680
01:03:47,560 --> 01:03:51,720
yeah open source is having a bit of a good time at the moment they're seeing some really

681
01:03:51,720 --> 01:03:59,680
impressive results there's money being invested into the space and obviously with big names

682
01:03:59,680 --> 01:04:07,360
like Meta contributing to the field it's definitely a place that businesses that want to deploy

683
01:04:07,360 --> 01:04:14,560
their own models will want to watch that there is the critical factor there companies that

684
01:04:14,560 --> 01:04:23,720
want to deploy their own models the use cases for these at the moment for many organisations

685
01:04:23,720 --> 01:04:32,480
it just isn't there so these feel like emergent there'll be some interesting developments

686
01:04:32,480 --> 01:04:39,760
in the coming months and years I'm sure but I think lots of organisations will first and

687
01:04:39,760 --> 01:04:51,880
foremost go through the APIs of models like Claw like OpenAI models because they've got

688
01:04:51,880 --> 01:04:56,800
the power and the heft and they can do what they want out of the box without having to

689
01:04:56,800 --> 01:05:06,960
do more technical jiggery-pokery yeah this is yeah I mean this is kind of yeah this is

690
01:05:06,960 --> 01:05:14,160
mostly still developer interest than business interest isn't it but A will start to see

691
01:05:14,160 --> 01:05:19,480
capabilities emerge that are easier to achieve with the open source models that are hard

692
01:05:19,480 --> 01:05:26,680
to achieve probably with the likes of Gemini and GPT-4 and it's just further evidence of

693
01:05:26,680 --> 01:05:30,840
the amount of innovation that's happening in many areas of models with now Google in

694
01:05:30,840 --> 01:05:34,560
the game Meta's got its new image model that have come out the video models are improving

695
01:05:34,560 --> 01:05:39,480
quickly we've got open source models chasing all of the commercial models and keeping them

696
01:05:39,480 --> 01:05:45,120
honest and keeping them developing their technologies so everything is continuing to race ahead

697
01:05:45,120 --> 01:05:51,320
at speed and the irony was we were going to make a joke about those of you that are waiting

698
01:05:51,320 --> 01:05:57,760
to sign up for chat GPT plus and still waiting for Gemini Ultra to come out that maybe the

699
01:05:57,760 --> 01:06:02,560
best model you could get your hands on was actually a free open source model for Mistral

700
01:06:02,560 --> 01:06:07,160
but for those of you who've been trying to get chat GPT plus and didn't see the news

701
01:06:07,160 --> 01:06:12,360
this week they opened up subscriptions again as of the 13th of December so if you've been

702
01:06:12,360 --> 01:06:19,080
waiting to get one of those now is the time to go get it and with that Christmas present

703
01:06:19,080 --> 01:06:22,440
for those of you that wanted that model and just haven't been able to get hold of it I

704
01:06:22,440 --> 01:06:29,280
think we shall say have a wonderful Christmas break we're probably going to do a special

705
01:06:29,280 --> 01:06:34,760
edition episode just before Christmas but we'll get back to reporting the news and having

706
01:06:34,760 --> 01:06:40,480
all of the interviews that you've known come to know and love starting at the beginning

707
01:06:40,480 --> 01:06:46,800
of the year 2024 next year have a fantastic Christmas break Martin.

708
01:06:46,800 --> 01:06:49,720
Same to you.

709
01:06:49,720 --> 01:06:53,720
And I will look forward to speaking to you soon.

710
01:06:53,720 --> 01:06:56,400
Merry Christmas everyone.

711
01:06:56,400 --> 01:07:02,000
Thank you for listening to artificially intelligent marketing to stay on top of the latest trends

712
01:07:02,000 --> 01:07:05,920
tips and tools in the world of marketing AI.

713
01:07:05,920 --> 01:07:07,680
Be sure to subscribe.

714
01:07:07,680 --> 01:07:11,240
We look forward to seeing you again next week.