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 --> 00:00:22,160
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:22,160 --> 00:00:23,920
Hello, everybody.

6
00:00:23,920 --> 00:00:29,200
Welcome to episode 38 of Artificially Intelligent Marketing and a Happy New Year to you all.

7
00:00:29,200 --> 00:00:33,200
It's the first episode of 2024.

8
00:00:33,200 --> 00:00:38,240
And as always, I'm joined by the wonderful, the wondrous, the magnificent Martin Broadhurst.

9
00:00:38,240 --> 00:00:40,040
How are you, Martin?

10
00:00:40,040 --> 00:00:41,040
I'm very good.

11
00:00:41,040 --> 00:00:50,280
2024 is upon us and I'm excited to see what it brings in terms of AI developments and

12
00:00:50,280 --> 00:00:53,600
Derby County successes, because we've not seen a lot of them in recent years.

13
00:00:53,600 --> 00:00:55,360
And I think this is the year, you know.

14
00:00:55,360 --> 00:01:00,960
I think you've got more chance of AGI this year than Derby County successes.

15
00:01:00,960 --> 00:01:03,160
And I'm happy to go on the record with that.

16
00:01:03,160 --> 00:01:06,600
In terms of AI developments, I can't tell you what's going to happen this year.

17
00:01:06,600 --> 00:01:09,560
I'm not sure anybody really knows, but I can tell you what's happened over the last couple

18
00:01:09,560 --> 00:01:13,680
of weeks and we can conjecture on where we might be heading, at least in the short term,

19
00:01:13,680 --> 00:01:15,160
because we've got some great stories this week.

20
00:01:15,160 --> 00:01:19,120
And we're going to be talking about OpenAI's GPT store going live.

21
00:01:19,120 --> 00:01:20,960
We're going to talk about chat GPT teams.

22
00:01:20,960 --> 00:01:23,960
We're going to talk about mid-journey version six.

23
00:01:23,960 --> 00:01:28,520
Talk about an interesting new tool called Bland AI, which is like a real-time AI phone

24
00:01:28,520 --> 00:01:30,360
call system, which is interesting.

25
00:01:30,360 --> 00:01:32,080
We're going to talk a little bit of robotics.

26
00:01:32,080 --> 00:01:35,120
There's been some cool robotics news this week.

27
00:01:35,120 --> 00:01:37,760
Martin and I are going to debate rag chatbots.

28
00:01:37,760 --> 00:01:38,760
Why aren't there more of them?

29
00:01:38,760 --> 00:01:40,680
If you don't know what a rag chatbot is, don't worry.

30
00:01:40,680 --> 00:01:43,240
We'll go through that in a bit more detail later on.

31
00:01:43,240 --> 00:01:46,520
We're going to talk perplexity, which is a very interesting tool that kind of is like

32
00:01:46,520 --> 00:01:49,760
a potential Google killer, but we'll have a chat about that as well.

33
00:01:49,760 --> 00:01:53,160
A little bit of conversation about New York Times suing OpenAI.

34
00:01:53,160 --> 00:01:57,480
Talk about what's been going on at CES 2024 and probably a whole host of other things

35
00:01:57,480 --> 00:01:58,960
in between.

36
00:01:58,960 --> 00:02:03,800
But to get us started, Martin, why don't we get into the GPT store?

37
00:02:03,800 --> 00:02:07,200
What is this and why is this important for the folks, the marketing and sales folks that

38
00:02:07,200 --> 00:02:08,680
might be listening to us today?

39
00:02:08,680 --> 00:02:14,120
Well, the big development that came out of OpenAI's developer day conference a couple

40
00:02:14,120 --> 00:02:16,680
of months ago was GPTs.

41
00:02:16,680 --> 00:02:24,160
And these were really, you could see them as kind of the next evolution of what had

42
00:02:24,160 --> 00:02:27,040
previously been plugins in chat GPT.

43
00:02:27,040 --> 00:02:36,160
And it's basically where you can connect external tools to chat GPT and give certain system

44
00:02:36,160 --> 00:02:40,960
prompts to give chat GPT certain characteristics or capabilities.

45
00:02:40,960 --> 00:02:49,560
And GPT or chat GPT plus subscribers have access to create their own GPTs for their

46
00:02:49,560 --> 00:02:52,960
own personal use, but you can share them publicly.

47
00:02:52,960 --> 00:02:58,680
And they announced at the developer day conference that the store was going to be launched later

48
00:02:58,680 --> 00:03:06,880
in the year where anybody can go in and use these publicly available GPTs.

49
00:03:06,880 --> 00:03:11,100
And the interesting thing was revenue share was included in that.

50
00:03:11,100 --> 00:03:20,040
So OpenAI basically said, if your GPT is used a lot by people, you will get a share of the

51
00:03:20,040 --> 00:03:24,880
revenues from the GPT plus subscriptions.

52
00:03:24,880 --> 00:03:27,480
And the store is now live.

53
00:03:27,480 --> 00:03:34,480
So if anybody's interested in using it, you need a chat GPT plus subscription, go on the

54
00:03:34,480 --> 00:03:42,640
left hand side and click on explore GPTs where it will bring up a panel in the main area

55
00:03:42,640 --> 00:03:49,320
where you can search and then click on any GPTs that you want.

56
00:03:49,320 --> 00:03:59,160
It's been widely adopted and one user who has got a lot of GPTs in the top 10 of the

57
00:03:59,160 --> 00:04:06,320
various different categories, because they're all categorized by a use case really.

58
00:04:06,320 --> 00:04:16,200
He has seen a hundred X increase in traffic to his website since the launch of the store.

59
00:04:16,200 --> 00:04:22,360
So there's clearly lots of users and what we're seeing is there's a lot of eyeballs

60
00:04:22,360 --> 00:04:32,240
to be captured if you have a top ranking GPT in any of the categories.

61
00:04:32,240 --> 00:04:36,560
It's quite clever actually, some of that, because having been playing with some of those,

62
00:04:36,560 --> 00:04:42,560
some of the more powerful GPTs, a lot of them have found quite clever ways to promote their

63
00:04:42,560 --> 00:04:44,880
own offering as part of the output.

64
00:04:44,880 --> 00:04:51,280
So an example would be an SEO GPT that maybe helps you do on page optimization for a given

65
00:04:51,280 --> 00:04:55,680
page where you'll get really good advice out of the GPT as you would expect.

66
00:04:55,680 --> 00:04:59,800
And then the last output or the last bullet point, so any advice is if you really want

67
00:04:59,800 --> 00:05:06,620
to dive deeper into this, check out XYZ website where obviously that website is the professional

68
00:05:06,620 --> 00:05:14,040
website of the person who made the GPT who just so happens to offer SEO consultancy capabilities.

69
00:05:14,040 --> 00:05:18,100
So I think people are finding quite clever ways to promote themselves in their GPTs outputs

70
00:05:18,100 --> 00:05:23,240
and that's even before we have details from OpenAI and what that revenue sharing model

71
00:05:23,240 --> 00:05:24,240
is going to be.

72
00:05:24,240 --> 00:05:29,320
So I'm really impressed with the cleverness of people to figure out ways to weave their

73
00:05:29,320 --> 00:05:34,320
own brands into these GPTs and use them as brand awareness and potentially even website

74
00:05:34,320 --> 00:05:38,440
traffic generating tools before they even get paid to even have them there.

75
00:05:38,440 --> 00:05:43,840
Yeah, well there's a will, there's a way and marketers are usually the first to take

76
00:05:43,840 --> 00:05:46,920
over a new channel, right?

77
00:05:46,920 --> 00:05:51,080
If we can take over and spam it with our brands, we will.

78
00:05:51,080 --> 00:05:52,400
Unfortunately so.

79
00:05:52,400 --> 00:05:57,680
I did see a couple of LinkedIn posts where people were warning that some of the GPTs

80
00:05:57,680 --> 00:06:01,320
seem engineered up to try and get you to give them your personal details.

81
00:06:01,320 --> 00:06:06,200
So another public service announcement here as we often do, don't go putting any sensitive

82
00:06:06,200 --> 00:06:12,940
details into a GPT, especially if it's one that appears to send data to another place.

83
00:06:12,940 --> 00:06:17,760
Maybe when you're using a GPT, it will tell you when it needs to send data to another

84
00:06:17,760 --> 00:06:22,000
place via API and you'll actually have to approve that.

85
00:06:22,000 --> 00:06:26,400
But yes, that's kind of an extra layer of self-protection is if you're using a GPT and

86
00:06:26,400 --> 00:06:30,280
it connects to a third party service, just be careful about what you're putting in and

87
00:06:30,280 --> 00:06:35,320
think twice before you allow it to because in the same way marketers get involved very

88
00:06:35,320 --> 00:06:39,240
quickly and potentially ruin channels, you've also got people who figure out a quick way

89
00:06:39,240 --> 00:06:43,360
to steal your personal identification and personal information and find a way to start

90
00:06:43,360 --> 00:06:44,400
making money out of that.

91
00:06:44,400 --> 00:06:46,800
So interesting platform.

92
00:06:46,800 --> 00:06:51,840
I've been playing with quite a few GPTs that I think are quite cool, but I already can

93
00:06:51,840 --> 00:06:55,640
see potential abuses left, right and center that we're all going to have to be very careful

94
00:06:55,640 --> 00:06:56,640
about.

95
00:06:56,640 --> 00:07:02,600
Any particular GPTs that you've found useful or fun?

96
00:07:02,600 --> 00:07:07,600
So do you know what I've trying to get into a habit of before I go to chat GPT and start

97
00:07:07,600 --> 00:07:10,560
a conversation to get something done?

98
00:07:10,560 --> 00:07:14,720
Searching the GPT store to see if there's a GPT for that that might have just been primed

99
00:07:14,720 --> 00:07:19,360
in some way to enable it to generate better outputs.

100
00:07:19,360 --> 00:07:24,200
So I'm still kind of in love with the first GPTs that I found because people were sharing

101
00:07:24,200 --> 00:07:28,640
them on Twitter and LinkedIn and providing public links like the convert anything GPT

102
00:07:28,640 --> 00:07:29,640
is still pretty cool.

103
00:07:29,640 --> 00:07:34,600
Give it an image of one sort and converts it into another like PNG to JPEG, convert

104
00:07:34,600 --> 00:07:37,720
audio from different formats, video from different formats.

105
00:07:37,720 --> 00:07:42,240
That one's pretty cool.

106
00:07:42,240 --> 00:07:48,040
I'm still on the fence a bit as to how good is it to have a GPT versus just have a conversation

107
00:07:48,040 --> 00:07:53,800
with chat GPT and I'm still to find a GPT where I'm like, oh wow, I could not have done

108
00:07:53,800 --> 00:07:56,840
this in chat GPT alone.

109
00:07:56,840 --> 00:08:00,040
Yeah, the power is going to come from the function calling, right?

110
00:08:00,040 --> 00:08:05,120
The connected GPTs to third party systems and external software.

111
00:08:05,120 --> 00:08:06,120
Yeah.

112
00:08:06,120 --> 00:08:09,680
And when I tried to build those myself, it was really ropey and broke really easily.

113
00:08:09,680 --> 00:08:13,800
So I suspect a lot of people have been having those problems.

114
00:08:13,800 --> 00:08:17,280
And a lot of the GPTs I play with don't do much function calling.

115
00:08:17,280 --> 00:08:19,720
Even the convert anything doesn't seem to do much.

116
00:08:19,720 --> 00:08:24,120
I think it's using like the Python libraries to actually, we're getting a bit detailed

117
00:08:24,120 --> 00:08:26,240
now, but I don't think it's sending information out.

118
00:08:26,240 --> 00:08:30,320
I think it's using what's already built into chat GPT to enable those conversions and being

119
00:08:30,320 --> 00:08:36,080
clever with code interpreter and our advanced data analysis to enable some of that stuff.

120
00:08:36,080 --> 00:08:39,360
But yeah, I think it will evolve.

121
00:08:39,360 --> 00:08:45,040
It's so much easier to use than the plugins that from a UX perspective, you talk a lot

122
00:08:45,040 --> 00:08:49,960
about UX and have over the last year, it's got much more chance of catching on than plugins

123
00:08:49,960 --> 00:08:52,720
because there's not a complete pain in the ass to use.

124
00:08:52,720 --> 00:08:56,840
Yeah, plugins were just a nightmare and never quite delivered on the promise.

125
00:08:56,840 --> 00:09:01,720
I don't think I ever really embraced them, but I see more potential for that with the

126
00:09:01,720 --> 00:09:03,520
GPT store.

127
00:09:03,520 --> 00:09:09,040
A moment ago, you gave a public service announcement talking about be careful with what data you

128
00:09:09,040 --> 00:09:10,600
put into chat GPT.

129
00:09:10,600 --> 00:09:16,680
Well we've had another development in that domain with OpenAI this week where they announced

130
00:09:16,680 --> 00:09:20,760
chat GPT for Teams that is now live.

131
00:09:20,760 --> 00:09:28,120
So previously, if you were an enterprise, you were a business, you had to use chat GPT

132
00:09:28,120 --> 00:09:35,240
enterprise which would, well first of all it came with quite a fee as far as we understand,

133
00:09:35,240 --> 00:09:39,240
but all of the data that you put into it was kept secure.

134
00:09:39,240 --> 00:09:46,880
Unlike if you're a free chat GPT or chat GPT plus subscriber whose data will be used to

135
00:09:46,880 --> 00:09:50,320
train the models in the future.

136
00:09:50,320 --> 00:10:00,120
Well now, OpenAI have announced chat GPT for Teams where the data stays safe.

137
00:10:00,120 --> 00:10:08,040
This is a $30 a month subscription or $25 per month if you pay annually.

138
00:10:08,040 --> 00:10:11,640
There's a minimum subscription so you need at least two people on the subscription.

139
00:10:11,640 --> 00:10:12,640
Makes sense.

140
00:10:12,640 --> 00:10:15,120
It's a Teams license, right?

141
00:10:15,120 --> 00:10:19,960
But yeah, this launch this week you get a higher usage cap so you can use GPT for a

142
00:10:19,960 --> 00:10:24,000
higher rate than you can on chat GPT plus.

143
00:10:24,000 --> 00:10:29,360
You have access to the 32k context window and your data is secure.

144
00:10:29,360 --> 00:10:31,280
So yeah, what do you think of that Paul?

145
00:10:31,280 --> 00:10:32,760
I'm pretty excited.

146
00:10:32,760 --> 00:10:37,680
We've talked about this in different ways and have been forced is the wrong word, but

147
00:10:37,680 --> 00:10:41,520
we've had to go looking for tools that we can use where we feel confident that we can

148
00:10:41,520 --> 00:10:47,680
put our own company's data in, which in some cases is a bit of a pain, especially because

149
00:10:47,680 --> 00:10:54,480
chat GPT's capabilities, once they won't be on text, code interpreter, stroke data analysis,

150
00:10:54,480 --> 00:10:57,400
give it an image and it can see what's in the image and all those other cool things

151
00:10:57,400 --> 00:10:58,560
you could do.

152
00:10:58,560 --> 00:11:03,720
They made it a bit harder to just use a text based generating tool because there's cool

153
00:11:03,720 --> 00:11:05,960
stuff that you want to do in chat GPT that you can't.

154
00:11:05,960 --> 00:11:10,240
We run our workshops, Martin, where we throw a load of dummy sales data in and then show

155
00:11:10,240 --> 00:11:15,080
how you could use that to do analysis for your sales data or even automatically create

156
00:11:15,080 --> 00:11:19,640
reports exported as PowerPoint files for you to share with internal team members, which

157
00:11:19,640 --> 00:11:23,120
is all well and good, but it's limited by the fact that your demonstration is a bunch

158
00:11:23,120 --> 00:11:26,920
of dummy data and then people are like, yeah, can't wait to get started with this.

159
00:11:26,920 --> 00:11:29,760
Where's all my proprietary unique highly sensitive sales data?

160
00:11:29,760 --> 00:11:32,480
And we're like, whoa, don't put it in there.

161
00:11:32,480 --> 00:11:33,760
You can't.

162
00:11:33,760 --> 00:11:39,480
And now if we understand everything correctly from chat GPT and the team at OpenAI, now you

163
00:11:39,480 --> 00:11:40,480
can.

164
00:11:40,480 --> 00:11:41,480
So I'm very excited about that.

165
00:11:41,480 --> 00:11:43,040
You and I have been waiting ages for enterprise.

166
00:11:43,040 --> 00:11:46,080
I've been on the waiting list for enterprise forever.

167
00:11:46,080 --> 00:11:54,320
Bios Droughta with its 20 employees was very low on the enterprise priority list for OpenAI.

168
00:11:54,320 --> 00:11:57,840
So when I've spoken about this, when I've been out and about at different conferences

169
00:11:57,840 --> 00:12:03,540
or speaking on AI and marketing, that's been the best I can recommend to people.

170
00:12:03,540 --> 00:12:05,480
Get on the waiting list for enterprise.

171
00:12:05,480 --> 00:12:08,240
And if you're a big enough organization, hurray, you might be able to get it.

172
00:12:08,240 --> 00:12:11,600
I'm so happy to see this bridge the gap.

173
00:12:11,600 --> 00:12:17,760
You can get chat GPT teams up to 149 users.

174
00:12:17,760 --> 00:12:21,160
So if you want everyone in your business to have it, then you have to be a team of under

175
00:12:21,160 --> 00:12:23,160
150 odd people.

176
00:12:23,160 --> 00:12:25,580
But not everyone in your business is going to need it.

177
00:12:25,580 --> 00:12:29,920
So I can imagine this could be appropriate for organizations up to as much as a thousand

178
00:12:29,920 --> 00:12:34,320
employees, you know, depending on what types of roles you have in your business.

179
00:12:34,320 --> 00:12:40,000
And it's pretty cost effective, really not a huge jump from the original chat GPT.

180
00:12:40,000 --> 00:12:48,200
One thing we talk about a lot is don't get coerced into buying an annual subscription

181
00:12:48,200 --> 00:12:50,440
because it seems a bit cheaper.

182
00:12:50,440 --> 00:12:55,880
We've got chat GPT teams at Bystrat already pretty much bought it as soon as it came out.

183
00:12:55,880 --> 00:13:00,800
It's tempting, but who knows like Gemini Ultra comes out in a couple of months time and then

184
00:13:00,800 --> 00:13:04,960
everyone's like, oh, it turns out Gemini Ultra from Google was actually better than GPT-4

185
00:13:04,960 --> 00:13:05,960
for all these use cases.

186
00:13:05,960 --> 00:13:06,960
And then you're like, oh, crumb.

187
00:13:06,960 --> 00:13:07,960
So I want to switch.

188
00:13:07,960 --> 00:13:10,200
But I've already paid for a year of this one.

189
00:13:10,200 --> 00:13:12,760
So I think the advice still holds mine.

190
00:13:12,760 --> 00:13:16,840
Don't pay for any annual subscriptions to save a few bucks because the new tools that

191
00:13:16,840 --> 00:13:20,400
are going to come out over the coming months could make that actually, it looks like a

192
00:13:20,400 --> 00:13:22,520
wise investment now, but I don't suspect it is.

193
00:13:22,520 --> 00:13:25,560
And you should only really be paying for monthly subscriptions.

194
00:13:25,560 --> 00:13:26,560
I think that's been our view.

195
00:13:26,560 --> 00:13:27,560
Do you agree with that, Ryan?

196
00:13:27,560 --> 00:13:30,560
Yeah, it's not what I practice, but yeah, it's my view.

197
00:13:30,560 --> 00:13:33,560
You've got to drink your own Kool-Aid, mate.

198
00:13:33,560 --> 00:13:34,560
Come on.

199
00:13:34,560 --> 00:13:38,200
No, I ended up going straight in on the annual.

200
00:13:38,200 --> 00:13:39,200
Yeah.

201
00:13:39,200 --> 00:13:46,280
The belief of that is I don't see myself moving away from chat GPT specifically.

202
00:13:46,280 --> 00:13:51,880
I actually agree with you very much so on other tools built on top of things.

203
00:13:51,880 --> 00:13:58,000
Things like HeyGen, things like 11labs, various of the tools I have subscriptions for, they're

204
00:13:58,000 --> 00:13:59,720
all monthlies.

205
00:13:59,720 --> 00:14:05,640
I don't see myself moving away from chat GPT within the next 12 months.

206
00:14:05,640 --> 00:14:06,640
I agree with you.

207
00:14:06,640 --> 00:14:07,640
I'll be honest.

208
00:14:07,640 --> 00:14:11,760
I talk about Magi a lot on this show and to be honest, moving to chat GPT teams might

209
00:14:11,760 --> 00:14:16,960
see us move away from Magi, but I also know Magi's developers are working to build multimodal

210
00:14:16,960 --> 00:14:18,520
capabilities into Magi.

211
00:14:18,520 --> 00:14:22,040
So when that comes out, I'm really not sure what tool I'm going to use because Magi are

212
00:14:22,040 --> 00:14:27,140
also this week, I think today, about to add Gemini Pro.

213
00:14:27,140 --> 00:14:32,600
So access to Google's models and chat GPT in the same tool for less than you'd pay for

214
00:14:32,600 --> 00:14:34,520
chat GPT.

215
00:14:34,520 --> 00:14:40,720
So I honestly use Magi more often than I use chat GPT unless I need a specific use case

216
00:14:40,720 --> 00:14:46,200
where I want to upload data or I want to upload images and have it analyze the images, which

217
00:14:46,200 --> 00:14:51,360
of course commercially I couldn't do because I couldn't trust the chat GPT account that

218
00:14:51,360 --> 00:14:52,360
I had.

219
00:14:52,360 --> 00:14:54,680
This actually puts the cat among the pigeons a little bit for me.

220
00:14:54,680 --> 00:14:59,160
And now I've got to see how Magi develops and responds, but that's why I still want

221
00:14:59,160 --> 00:15:00,320
the flexibility.

222
00:15:00,320 --> 00:15:06,320
We could probably spend many hours debating this, Martin, but I guess for the listeners,

223
00:15:06,320 --> 00:15:08,840
go check out chat GPT teams.

224
00:15:08,840 --> 00:15:14,200
If you go and have a look at the webpage for this, it's very interesting that lots of

225
00:15:14,200 --> 00:15:20,480
testimonials from customers and quotes and examples are clearly deliberately engineered

226
00:15:20,480 --> 00:15:23,640
to give you confidence that you can put your own business data in.

227
00:15:23,640 --> 00:15:29,000
One of them is like, help me write a blog post ahead of our new product launch.

228
00:15:29,000 --> 00:15:32,400
Here's all the unique features of that product, which of course you'd never do before because

229
00:15:32,400 --> 00:15:34,920
you didn't want to put sensitive proprietary information in.

230
00:15:34,920 --> 00:15:39,120
Another example is, oh, I upload my sales data and I analyze it a bit like the examples

231
00:15:39,120 --> 00:15:40,540
that we've given in the past.

232
00:15:40,540 --> 00:15:44,040
So they're clearly trying to give us all confidence that we can put sensitive information into

233
00:15:44,040 --> 00:15:47,020
chat GPT if we've got chat GPT teams.

234
00:15:47,020 --> 00:15:49,000
So as a user, go have a look, check it out.

235
00:15:49,000 --> 00:15:52,040
It might be that you can do cool stuff with it that you didn't feel comfortable to do

236
00:15:52,040 --> 00:15:54,520
before.

237
00:15:54,520 --> 00:15:56,640
Should we move on to our next story, Martino?

238
00:15:56,640 --> 00:15:57,640
Let's do it.

239
00:15:57,640 --> 00:15:58,640
Let's do it.

240
00:15:58,640 --> 00:16:00,800
Next one is an update to Mid Journey.

241
00:16:00,800 --> 00:16:05,560
So Mid Journey has been testing version six of its image generator for listeners to the

242
00:16:05,560 --> 00:16:08,960
podcast who are probably quite au fait with Mid Journey.

243
00:16:08,960 --> 00:16:13,480
But for those that are not, Mid Journey is one of those power tools for generating AI

244
00:16:13,480 --> 00:16:14,480
images.

245
00:16:14,480 --> 00:16:19,400
In our perspective, it generates the most photorealistic images of all the sets, especially

246
00:16:19,400 --> 00:16:23,720
because that even though Dually 3, which you can access through chat GPT can do it, it

247
00:16:23,720 --> 00:16:28,120
did get the feeling towards the end of last year that that photorealism was being watered

248
00:16:28,120 --> 00:16:29,120
down a bit.

249
00:16:29,120 --> 00:16:31,360
So Mid Journey is still your best for that.

250
00:16:31,360 --> 00:16:36,640
And people have been playing with it and getting it to do some pretty cool stuff.

251
00:16:36,640 --> 00:16:41,860
Although it does tend to have a tendency to almost over detail sort of, it's got an obsession

252
00:16:41,860 --> 00:16:44,800
with I don't know if you've been playing with this, Martin, but it's got an obsession

253
00:16:44,800 --> 00:16:46,720
with skin details.

254
00:16:46,720 --> 00:16:54,560
It's almost like everybody has to have freckles and lines in their skin and blemishes and

255
00:16:54,560 --> 00:16:55,560
imperfections.

256
00:16:55,560 --> 00:16:59,360
It's almost like that's been dialed up to 11 just to show that it can do it.

257
00:16:59,360 --> 00:17:03,080
But it's an alpha and I'm sure all of that stuff will get rebalanced.

258
00:17:03,080 --> 00:17:06,840
Perhaps the most important improvement is it can now do text.

259
00:17:06,840 --> 00:17:09,960
If you want to give it lots of text, it struggles, you're better with one or two words.

260
00:17:09,960 --> 00:17:14,320
But there's this ongoing arms race between Mid Journey and Dually 3, it feels like, and

261
00:17:14,320 --> 00:17:18,640
Mid Journey just threw out its next big weapon to show us what it's capable of.

262
00:17:18,640 --> 00:17:20,880
And I think it's got a lot of interesting stuff.

263
00:17:20,880 --> 00:17:24,800
Have you been paying attention to Mid Journey version 6 at all, Martin, how to play or anything

264
00:17:24,800 --> 00:17:26,160
like that?

265
00:17:26,160 --> 00:17:27,160
Not so much.

266
00:17:27,160 --> 00:17:33,960
I still use Mid Journey through Discord, but I'm using V5 predominantly.

267
00:17:33,960 --> 00:17:34,960
I have seen the outputs.

268
00:17:34,960 --> 00:17:35,960
They look great.

269
00:17:35,960 --> 00:17:38,200
I mean, this is the one thing that we can say for Mid Journey.

270
00:17:38,200 --> 00:17:41,360
The outputs consistently look fantastic.

271
00:17:41,360 --> 00:17:47,760
And actually, I do like them more than the OpenAI ones.

272
00:17:47,760 --> 00:17:55,920
OpenAI feels there's something about them now in Dually 3 where I see them and they

273
00:17:55,920 --> 00:17:58,880
just have an imprint of Dually 3, right?

274
00:17:58,880 --> 00:18:00,520
They have a look and feel about them.

275
00:18:00,520 --> 00:18:04,440
Unless you get really into being specific on the style, if you just say, oh, I want

276
00:18:04,440 --> 00:18:08,640
a picture like this, or I want an image of such and such, there's something about them

277
00:18:08,640 --> 00:18:10,640
that's immediately recognisable, I think.

278
00:18:10,640 --> 00:18:16,600
It's a bit like every time I drink a BrewDog beer, I'm like, this has a BrewDog beer characteristic

279
00:18:16,600 --> 00:18:17,600
to it.

280
00:18:17,600 --> 00:18:23,000
It's like every time you go to a Derby County performance and you come away sort of mildly

281
00:18:23,000 --> 00:18:27,960
disappointed you know before you go that that's what you're going to get.

282
00:18:27,960 --> 00:18:31,680
In many respects, I like the consistency, right?

283
00:18:31,680 --> 00:18:36,320
At least you're like, yeah, they never let you down in a weird sort of way.

284
00:18:36,320 --> 00:18:43,280
Dually 3, that part of it disappoints me because when it first came out, you didn't have that.

285
00:18:43,280 --> 00:18:47,840
When you accessed it through Bing, so if you don't have a Dually, sorry, ChatGPT Plus account,

286
00:18:47,840 --> 00:18:51,320
you can actually access it through what has been called Copilot now.

287
00:18:51,320 --> 00:18:55,200
We'll talk about that later in another segment.

288
00:18:55,200 --> 00:18:58,960
You can get it to produce images that you can't, you can get Bing, straight Copilot

289
00:18:58,960 --> 00:19:05,520
to produce images that Dually 3 in ChatGPT, there's so many brand and product name problems

290
00:19:05,520 --> 00:19:07,840
making it confusing for people.

291
00:19:07,840 --> 00:19:11,600
They did this, gang, sorry, we didn't.

292
00:19:11,600 --> 00:19:17,920
But with Bing, straight Copilot, you can get it to produce images that Dually 3 never.

293
00:19:17,920 --> 00:19:22,480
My favorite way to test it is to get Deadpool doing different things.

294
00:19:22,480 --> 00:19:27,880
And at the beginning, I could test Dually 3 against Bing and Bing would show me Deadpool

295
00:19:27,880 --> 00:19:34,940
and Bing and Dually 3 would say, sorry, Deadpool's copyrighted work and we can't show you it.

296
00:19:34,940 --> 00:19:38,600
And the photorealism even at that stage was better in the Bing version.

297
00:19:38,600 --> 00:19:41,200
So I think they're watering Dually 3 down.

298
00:19:41,200 --> 00:19:46,040
I don't know if it's a compute thing or a copyright thing or a deepfake thing, but it

299
00:19:46,040 --> 00:19:49,480
does feel like they've done it deliberately to me.

300
00:19:49,480 --> 00:19:53,800
Some other image news this week was one of the biggest challenges.

301
00:19:53,800 --> 00:19:57,360
If you've been playing with Mid Journey, Dually 3, some of the tools we talk about on the

302
00:19:57,360 --> 00:20:01,280
podcast, you'll know that one of the awesome things would be to like create a comic book

303
00:20:01,280 --> 00:20:06,320
or something similar, right, or even like some storyboards.

304
00:20:06,320 --> 00:20:11,340
But one of the big challenges is having consistent characters in your outputs, which there are

305
00:20:11,340 --> 00:20:15,620
some workarounds you can use something called the seed value to try and ensure that the

306
00:20:15,620 --> 00:20:17,580
characters in your images stay similar.

307
00:20:17,580 --> 00:20:21,660
You could do some stuff with that, but it's very hard to have full control over what that

308
00:20:21,660 --> 00:20:22,660
character is doing.

309
00:20:22,660 --> 00:20:24,800
And it just basically doesn't really work.

310
00:20:24,800 --> 00:20:28,960
But there was some research that came out of Bike Dance, which is the parent company

311
00:20:28,960 --> 00:20:34,640
of TikTok, who have created a tool called DreamTuner, which they say makes it much easier

312
00:20:34,640 --> 00:20:37,480
to do subject driven generation from a single image.

313
00:20:37,480 --> 00:20:42,120
So basically you give it an image that sets the character, and then you can generate subsequent

314
00:20:42,120 --> 00:20:44,360
images where the character stays the same.

315
00:20:44,360 --> 00:20:49,480
So we won't get sort of into the details too much of how they do that.

316
00:20:49,480 --> 00:20:54,280
But it opens up the possibility that in the near future, the image generation tools that

317
00:20:54,280 --> 00:20:59,440
we're all using could potentially have the power to create consistent characters, which

318
00:20:59,440 --> 00:21:05,360
opens up a lot of creative avenues in marketing and outside of marketing that we can't currently

319
00:21:05,360 --> 00:21:07,440
do, but many users wish they could.

320
00:21:07,440 --> 00:21:13,000
So I think that's something to keep an eye on and something we've been talking about

321
00:21:13,000 --> 00:21:15,640
on the podcast before, Martin.

322
00:21:15,640 --> 00:21:19,700
The examples in the research paper, and it is still just a research paper, right?

323
00:21:19,700 --> 00:21:22,860
This is not in production or available for anyone to use yet.

324
00:21:22,860 --> 00:21:26,040
But the examples in the research paper do look really promising.

325
00:21:26,040 --> 00:21:31,520
So you can be sure that industry is all over this.

326
00:21:31,520 --> 00:21:34,280
People want to see this deployed as quickly as possible.

327
00:21:34,280 --> 00:21:39,760
So if I was going to make any prediction about what we might expect by the end of 2024, consistent

328
00:21:39,760 --> 00:21:43,840
character development via AI image generation might be one of the things I would look out

329
00:21:43,840 --> 00:21:44,840
for.

330
00:21:44,840 --> 00:21:45,840
I agree.

331
00:21:45,840 --> 00:21:49,960
I think they know exactly where the limitations are and they're just crossing them out.

332
00:21:49,960 --> 00:21:52,920
Like text generation, we all wanted text, it's like, oh, wouldn't that be great?

333
00:21:52,920 --> 00:21:53,920
And now we have it.

334
00:21:53,920 --> 00:21:56,520
And I think this is absolutely in the list that you're right, Martin.

335
00:21:56,520 --> 00:22:01,360
Was there some news as well about mid-journey and video generation that you saw?

336
00:22:01,360 --> 00:22:02,360
Was there something in there?

337
00:22:02,360 --> 00:22:06,080
Yeah, they're going to start training video models this month.

338
00:22:06,080 --> 00:22:13,340
It was discussed by the mid-journey CEO on a Q&A on their Discord channel.

339
00:22:13,340 --> 00:22:20,440
So when that actually appears for production and use or an alpha version remains to be

340
00:22:20,440 --> 00:22:26,980
seen, but they are moving into the AI video generation game.

341
00:22:26,980 --> 00:22:36,040
Not necessarily surprising given video is basically 24 images per second.

342
00:22:36,040 --> 00:22:42,240
I'm quite excited about this news because I see mid-journey as the highest quality images

343
00:22:42,240 --> 00:22:46,680
and a lot of things about video as the underlying image to begin with.

344
00:22:46,680 --> 00:22:51,060
If they are able to stitch those images together in a meaningful way and not get any of these

345
00:22:51,060 --> 00:22:57,260
crazy artifacts that we've seen in other tools like Runway and Pica, although they are improving

346
00:22:57,260 --> 00:23:02,920
at extreme speed, both of those tools, it could be quite another interesting player

347
00:23:02,920 --> 00:23:04,160
in the game.

348
00:23:04,160 --> 00:23:07,640
And as we talk a lot on the podcast, the more players trying to improve these tools, the

349
00:23:07,640 --> 00:23:09,240
faster we're going to see progress.

350
00:23:09,240 --> 00:23:13,760
So yeah, I'm excited to see what mid-journey come up with.

351
00:23:13,760 --> 00:23:16,120
Let's talk bland.ai Martin.

352
00:23:16,120 --> 00:23:19,160
This is quite an interesting tool that's been getting a bit of buzz over the last week or

353
00:23:19,160 --> 00:23:20,160
so.

354
00:23:20,160 --> 00:23:21,160
It is.

355
00:23:21,160 --> 00:23:26,960
And it might be the thing that kills off cold calling eternally for me.

356
00:23:26,960 --> 00:23:33,200
So this is a real time AI phone call system designed for sales, customer service and more.

357
00:23:33,200 --> 00:23:37,320
And yeah, it's been making waves, shall we say.

358
00:23:37,320 --> 00:23:43,560
Someone said that their jaw was on the floor after testing it, such was its capabilities.

359
00:23:43,560 --> 00:23:50,080
So it's capable of sending up to 500,000 calls, which is just what?

360
00:23:50,080 --> 00:23:51,080
Yeah.

361
00:23:51,080 --> 00:23:53,480
Unfathomably large.

362
00:23:53,480 --> 00:24:00,680
So it allows you to scale up AI sales calls, right?

363
00:24:00,680 --> 00:24:09,160
Some people have said that their voices sometimes sound a little bland, but you're doing AI

364
00:24:09,160 --> 00:24:11,360
sales calls at scale.

365
00:24:11,360 --> 00:24:16,700
And the reason I say that, I think this is potentially a cold call killer is if everybody

366
00:24:16,700 --> 00:24:20,480
has this power to just suddenly, it's like cold outreach.

367
00:24:20,480 --> 00:24:25,720
If I ask you about your inbox now, Paul, and say in your inbox, have you got any emails

368
00:24:25,720 --> 00:24:31,680
that you think are sent via a sales automation platform as part of a sales sequence?

369
00:24:31,680 --> 00:24:33,960
I imagine you've got a couple in there.

370
00:24:33,960 --> 00:24:34,960
Yeah.

371
00:24:34,960 --> 00:24:37,000
Far too many for my liking.

372
00:24:37,000 --> 00:24:38,000
Yeah.

373
00:24:38,000 --> 00:24:41,560
And this feels like the same.

374
00:24:41,560 --> 00:24:44,280
I can see the use cases in customer service though.

375
00:24:44,280 --> 00:24:49,960
If it's incoming calls and you've got something that's a much more lifelike voice, much more

376
00:24:49,960 --> 00:24:53,600
almost character AI driven conversation.

377
00:24:53,600 --> 00:24:56,880
I think that will work very well.

378
00:24:56,880 --> 00:25:03,600
In terms of sales activity, if it was cold, I think that's a terrifying prospect if it

379
00:25:03,600 --> 00:25:07,240
can be rolled out at scale at this level.

380
00:25:07,240 --> 00:25:08,240
Yeah.

381
00:25:08,240 --> 00:25:10,200
That bit makes me cry a bit.

382
00:25:10,200 --> 00:25:12,960
I really don't want that.

383
00:25:12,960 --> 00:25:14,080
But I think you're right.

384
00:25:14,080 --> 00:25:18,520
The customer service part has some power to it.

385
00:25:18,520 --> 00:25:22,160
Because I think the key here is with great power comes great responsibility.

386
00:25:22,160 --> 00:25:25,080
And we know as we've talked about on this podcast already, that probably is going to

387
00:25:25,080 --> 00:25:27,720
be misused and that is the worry.

388
00:25:27,720 --> 00:25:32,280
But if you can in your organization find a way to leverage tools like these that benefit

389
00:25:32,280 --> 00:25:37,680
the customer and you do it well, well you're increasing response times to customer service

390
00:25:37,680 --> 00:25:38,680
calls.

391
00:25:38,680 --> 00:25:43,360
If you can ensure the quality of the advice that's given on those calls and people no

392
00:25:43,360 --> 00:25:47,720
longer have to wait in lines to speak to someone, well you're going to be adding value to customers

393
00:25:47,720 --> 00:25:53,240
and they're probably going to be willing to speak with a robot even though they'd probably

394
00:25:53,240 --> 00:25:54,560
rather speak with a human.

395
00:25:54,560 --> 00:25:57,440
Because if you can get them to their solution faster, they're going to appreciate it.

396
00:25:57,440 --> 00:26:02,240
What they're not going to appreciate is the type of robo cold calls that you mentioned

397
00:26:02,240 --> 00:26:03,240
Martin.

398
00:26:03,240 --> 00:26:07,360
I'd be wrong to say I've been playing with it, but I've been watching some of the demos

399
00:26:07,360 --> 00:26:11,080
and sort of partaking in some of the conversation online.

400
00:26:11,080 --> 00:26:16,680
What this reminds me of is Google's demo many years ago where they showed us in essence

401
00:26:16,680 --> 00:26:22,040
an AI that would make calls on your behalf, book restaurants for you, would know the right

402
00:26:22,040 --> 00:26:29,600
questions to ask, had like little human fake ums and ahs and pauses.

403
00:26:29,600 --> 00:26:33,980
I think the reason it's making buzz is because this is the first genuine product that anybody

404
00:26:33,980 --> 00:26:35,600
can access that can do that.

405
00:26:35,600 --> 00:26:42,960
But you're right, some of the critical feedback has been that AI responds quick, which is

406
00:26:42,960 --> 00:26:49,040
great because they know a large language model needs to process what the person said and

407
00:26:49,040 --> 00:26:50,380
then come up with a response.

408
00:26:50,380 --> 00:26:54,380
They know response time is important, but it seems to me that they've optimized it

409
00:26:54,380 --> 00:26:58,720
so much that the AI always speaks back too quickly, whereas most humans have to pause

410
00:26:58,720 --> 00:27:01,800
for a second to think about what they've heard.

411
00:27:01,800 --> 00:27:08,380
The other criticism was that there wasn't enough natural variation in how the fake person

412
00:27:08,380 --> 00:27:12,640
in this case speaks to be believable.

413
00:27:12,640 --> 00:27:18,600
We'd all learn quite quickly to be able to differentiate what was a real person and what

414
00:27:18,600 --> 00:27:22,760
wasn't because these tools are not organic enough yet.

415
00:27:22,760 --> 00:27:24,080
I'm sure they will be.

416
00:27:24,080 --> 00:27:26,120
This is like the first one of these that we're seeing.

417
00:27:26,120 --> 00:27:27,120
So interesting.

418
00:27:27,120 --> 00:27:34,360
Also, just talking about the AI capabilities of dealing with customer service queries,

419
00:27:34,360 --> 00:27:39,920
I do wonder how much of this is going to be done by voice and how much of it will be just

420
00:27:39,920 --> 00:27:43,400
done via text interfaces.

421
00:27:43,400 --> 00:27:44,400
Maybe voice.

422
00:27:44,400 --> 00:27:45,400
I don't know.

423
00:27:45,400 --> 00:27:51,160
I'm just thinking about how much I use chat GPT voice and have the conversations with

424
00:27:51,160 --> 00:27:52,160
chat GPT.

425
00:27:52,160 --> 00:27:57,160
And actually in real time, I'm having the argument in my head telling me that I'm wrong

426
00:27:57,160 --> 00:28:01,280
because there's an example with Six Flags.

427
00:28:01,280 --> 00:28:03,440
I don't know if you've seen this on the Google website.

428
00:28:03,440 --> 00:28:08,360
They've got a case study with the theme park Six Flags and how they're using generative

429
00:28:08,360 --> 00:28:09,440
AI.

430
00:28:09,440 --> 00:28:14,920
So they've connected this with the Google Cloud Vertex services and they've basically

431
00:28:14,920 --> 00:28:26,800
created an app with generative AI plugged into it and they're using RAG.

432
00:28:26,800 --> 00:28:33,360
So the retrieval system connected to their own knowledge system or knowledge base.

433
00:28:33,360 --> 00:28:37,880
And it's now enabling customers and visitors to their parks to ask questions, to create

434
00:28:37,880 --> 00:28:41,560
personalized itineraries for them on their visit.

435
00:28:41,560 --> 00:28:45,240
And if they've got questions about, you know, what time does this restaurant open and where

436
00:28:45,240 --> 00:28:51,560
can I find accessible parking or what have you, it will answer it in the app.

437
00:28:51,560 --> 00:28:55,400
And that's how I can see people.

438
00:28:55,400 --> 00:28:58,880
And I was imagining that as more of a text-based interface and maybe we're just going to move

439
00:28:58,880 --> 00:29:03,600
away from spoken interactions.

440
00:29:03,600 --> 00:29:10,400
But then I just thought about how much I enjoy using chat GPT voice and completely shot myself

441
00:29:10,400 --> 00:29:11,400
down.

442
00:29:11,400 --> 00:29:16,520
I think the reason it's hard to imagine, and I thought about this a lot, is because I think

443
00:29:16,520 --> 00:29:17,520
it's both.

444
00:29:17,520 --> 00:29:21,720
And I think it comes down to information transfer speed.

445
00:29:21,720 --> 00:29:27,040
So I'm at the point now where I'm dictating 50 to 80% of my emails, depending on the sort

446
00:29:27,040 --> 00:29:30,800
of content, how well I know what I want to say, basically.

447
00:29:30,800 --> 00:29:35,400
Like if I know what I want to say, I can just speak really fast, transcribes, it's in the

448
00:29:35,400 --> 00:29:37,840
email, I can make a few edits and away I go.

449
00:29:37,840 --> 00:29:41,560
If I don't know what I want to say, it's easier to write it in some cases, to be honest, because

450
00:29:41,560 --> 00:29:43,760
then I can reformulate it as I go.

451
00:29:43,760 --> 00:29:49,240
But certainly in general, giving information for me is much faster to speak than it is

452
00:29:49,240 --> 00:29:50,240
to type.

453
00:29:50,240 --> 00:29:51,840
So I would rather speak.

454
00:29:51,840 --> 00:29:55,620
Absorbing information, I do much faster reading than I do listening.

455
00:29:55,620 --> 00:30:02,200
So my frustration with ChatGPT on the phone, because as I understand it, you can't dictate

456
00:30:02,200 --> 00:30:06,960
to the desktop app yet still, you can only do it on mobile, which is something for me

457
00:30:06,960 --> 00:30:11,280
they need to fix because I also want to speak to ChatGPT on desktop, not just on the phone.

458
00:30:11,280 --> 00:30:14,560
But I don't really like having a conversation with ChatGPT on the phone, because I want

459
00:30:14,560 --> 00:30:16,420
to speak to it because it's fast for me.

460
00:30:16,420 --> 00:30:19,360
And then I want to read what ChatGPT says back because it's fast for me.

461
00:30:19,360 --> 00:30:22,220
I don't want to listen to ChatGPT because it's not fast enough.

462
00:30:22,220 --> 00:30:27,720
So I think it will be a mixture, honestly, because what should go?

463
00:30:27,720 --> 00:30:33,400
If you're in the kitchen and you're cooking and you want to get a bit of quick recommendation,

464
00:30:33,400 --> 00:30:36,720
hey, would it be good to put this spice in this meal?

465
00:30:36,720 --> 00:30:39,400
You're not going to go over and read something or you're not going to get your phone out

466
00:30:39,400 --> 00:30:40,400
of your pocket, right?

467
00:30:40,400 --> 00:30:42,400
You're going to want it to speak back to you.

468
00:30:42,400 --> 00:30:45,880
But for me, when information speed is critical, I want to read, not listen.

469
00:30:45,880 --> 00:30:50,640
So I don't think reading and writing in that case is going away.

470
00:30:50,640 --> 00:30:52,960
I think it's just what's appropriate for the context.

471
00:30:52,960 --> 00:30:55,760
What's the use case?

472
00:30:55,760 --> 00:30:59,600
Another sort of interesting thing in this area is Martin and I have been like proper

473
00:30:59,600 --> 00:31:05,360
nerding out on robotics over the last week or two, because there's been some real cool

474
00:31:05,360 --> 00:31:07,880
stuff going on.

475
00:31:07,880 --> 00:31:14,080
And whilst this is probably else, well, it's AI driven stuff, it's not the usual AI stuff

476
00:31:14,080 --> 00:31:17,360
we talk about, which is why we think it's important to talk about.

477
00:31:17,360 --> 00:31:23,160
So this week we saw the figure01 robot, so Fig is a company that have emerged out of

478
00:31:23,160 --> 00:31:26,160
stealth over the last 12 months or so doing some really cool stuff.

479
00:31:26,160 --> 00:31:32,680
And what they did is they have a new landmark development has been brewing inside their

480
00:31:32,680 --> 00:31:38,840
organization quite literally, ha ha ha, because their figure01 robot, which is basically guided

481
00:31:38,840 --> 00:31:43,520
by neural networks, has mastered the art of making a coffee.

482
00:31:43,520 --> 00:31:48,440
And so this seems sort of probably quite a simple thing to be able to do.

483
00:31:48,440 --> 00:31:53,400
But the reason that it's so interesting is it hasn't been programmed in the traditional

484
00:31:53,400 --> 00:31:57,280
way in terms of the steps you need to go through to make a coffee.

485
00:31:57,280 --> 00:32:02,800
It was trained, as I understand it, on videos of humans making a coffee.

486
00:32:02,800 --> 00:32:07,440
And then it learned from that information how to do it.

487
00:32:07,440 --> 00:32:12,120
This for me is an absolute critical game changer, because the hardest thing about programming

488
00:32:12,120 --> 00:32:18,080
robots is programming them in the traditional style of all the steps they need to do to

489
00:32:18,080 --> 00:32:19,720
do a certain thing.

490
00:32:19,720 --> 00:32:23,040
And then all the edge cases that can get in the way and cause issues.

491
00:32:23,040 --> 00:32:27,140
And when you operate in the real 3D world that we all live in, everything's an edge

492
00:32:27,140 --> 00:32:28,140
case, right?

493
00:32:28,140 --> 00:32:29,420
Like how big is the table?

494
00:32:29,420 --> 00:32:30,420
How tall is the table?

495
00:32:30,420 --> 00:32:31,960
What coffee machine is it?

496
00:32:31,960 --> 00:32:32,960
What size are the pods?

497
00:32:32,960 --> 00:32:34,680
What do I do if I drop a pod?

498
00:32:34,680 --> 00:32:37,480
What happens if the machine doesn't work the first time I press the button?

499
00:32:37,480 --> 00:32:39,540
Like it's a nightmare to code.

500
00:32:39,540 --> 00:32:43,620
So the idea that you could just give a load of video and have the robot learn from the

501
00:32:43,620 --> 00:32:45,420
video is critical.

502
00:32:45,420 --> 00:32:50,080
And this is also, as I understand it, what Tesla's now doing when it comes to its automatic

503
00:32:50,080 --> 00:32:51,080
driving.

504
00:32:51,080 --> 00:32:54,280
So there's a great book on Elon Musk.

505
00:32:54,280 --> 00:32:57,640
I can't remember the author, which is very naughty of me.

506
00:32:57,640 --> 00:33:01,920
But it's a very recent book out that's very worth reading.

507
00:33:01,920 --> 00:33:05,880
And the last parts of that book are in 2023.

508
00:33:05,880 --> 00:33:11,220
And they talk about the move that Tesla's making from programming how cars should react

509
00:33:11,220 --> 00:33:20,000
in different scenarios to just using the billions of frames of video data they're collecting

510
00:33:20,000 --> 00:33:25,680
from all of their cars to basically have the neural net teach itself how to drive in almost

511
00:33:25,680 --> 00:33:28,680
all conditions by just observing how humans did it.

512
00:33:28,680 --> 00:33:32,320
Which if I'm honest, is what I thought they were doing from the start, but it wasn't.

513
00:33:32,320 --> 00:33:34,280
And it is now what they're doing.

514
00:33:34,280 --> 00:33:38,120
So this is really interesting because you could imagine now training a robot to pretty

515
00:33:38,120 --> 00:33:43,320
much do anything if you had video data of humans doing it, and then give them the tasks.

516
00:33:43,320 --> 00:33:49,560
If they get the robotic aspects of it right, they'll be able to do anything.

517
00:33:49,560 --> 00:33:53,760
So robotics has been one of those things that ever since you saw Will Smith in iRobot, you're

518
00:33:53,760 --> 00:33:57,440
like, appreciate that all the robots were going to try and kill us, and that's not ideal.

519
00:33:57,440 --> 00:34:00,960
But cool, wouldn't it be nice to have robots to help us with a load of stuff?

520
00:34:00,960 --> 00:34:02,840
This is like a huge amount of progress for me.

521
00:34:02,840 --> 00:34:07,280
And I'm sure people inside this industry probably knew a bit more about how some of this stuff

522
00:34:07,280 --> 00:34:08,280
was coming.

523
00:34:08,280 --> 00:34:12,000
But the fact that you could get a robot to do things for you is kind of not that far

524
00:34:12,000 --> 00:34:13,000
off.

525
00:34:13,000 --> 00:34:14,000
That's how it feels anyway.

526
00:34:14,000 --> 00:34:15,000
Don't know what you think, Martin.

527
00:34:15,000 --> 00:34:18,760
I know you've been paying attention to Tesla's work in this area as well.

528
00:34:18,760 --> 00:34:28,000
Yeah, last month they announced Gen 2 of their humanoid robot Optimus, which again, is that

529
00:34:28,000 --> 00:34:35,320
when you said about getting the robot elements right, the actual mechanics of it, they're

530
00:34:35,320 --> 00:34:39,880
making sure that if it opens a door, it doesn't rip the door off its hinges, right?

531
00:34:39,880 --> 00:34:40,880
Things like that.

532
00:34:40,880 --> 00:34:43,080
These are really quite important.

533
00:34:43,080 --> 00:34:48,280
And they did a demo that showed this new humanoid robot, and it's got good balance and it walks

534
00:34:48,280 --> 00:34:50,360
faster than the previous version.

535
00:34:50,360 --> 00:34:51,360
There's really little things in it.

536
00:34:51,360 --> 00:34:54,800
Like the dexterity of the fingers is greatly improved.

537
00:34:54,800 --> 00:35:01,600
So they demonstrate that it can pick up an egg and it shows a visual of how the pressure

538
00:35:01,600 --> 00:35:08,640
pads on the fingers interpret the sensitivity and understand and interpret the, well, how

539
00:35:08,640 --> 00:35:11,120
it senses the world really.

540
00:35:11,120 --> 00:35:13,140
And yeah, it's just a huge development.

541
00:35:13,140 --> 00:35:17,400
And I think this is, it's happening very quickly.

542
00:35:17,400 --> 00:35:20,600
And Tesla clearly not the only game in town.

543
00:35:20,600 --> 00:35:31,280
They've got figure one, Boston Dynamics has had their machines out in, certainly in video

544
00:35:31,280 --> 00:35:35,560
form available to demo for a long time, whether they've actually gone into commercial applications

545
00:35:35,560 --> 00:35:38,840
yet, I'm not actually sure.

546
00:35:38,840 --> 00:35:40,880
But robotics is big.

547
00:35:40,880 --> 00:35:48,480
Amazon did a story at the back end of last year talking about how they deployed 750,000

548
00:35:48,480 --> 00:35:51,000
robots across their warehouses.

549
00:35:51,000 --> 00:35:56,080
And they were expecting to deploy that and more every year going forward.

550
00:35:56,080 --> 00:36:01,780
And they're also working on a humanoid robot to work in their factories alongside humans.

551
00:36:01,780 --> 00:36:06,680
The robots that they've got at the moment are all like production line type, little

552
00:36:06,680 --> 00:36:12,360
things to move boxes from A to B, but they're actually starting to look at deploying humanoid

553
00:36:12,360 --> 00:36:15,560
ones later in the year as well.

554
00:36:15,560 --> 00:36:20,120
So yeah, definitely something we all need to keep an eye on because before you know

555
00:36:20,120 --> 00:36:24,080
it, there'll be robots in every neighborhood.

556
00:36:24,080 --> 00:36:25,560
It's kind of interesting.

557
00:36:25,560 --> 00:36:32,600
I think that's where the developments will be felt most keenly.

558
00:36:32,600 --> 00:36:36,720
There are robots in picking factories for Amazon and other distributors, right?

559
00:36:36,720 --> 00:36:42,120
It's not like robotics is new to those environments, but the capabilities of those robots is improving

560
00:36:42,120 --> 00:36:43,520
quickly.

561
00:36:43,520 --> 00:36:50,800
I read a story a week or two ago about Samsung was planning human-free, fully automated fabs

562
00:36:50,800 --> 00:36:51,920
within six years.

563
00:36:51,920 --> 00:36:58,040
So basically trying to eliminate the need for any human workers by having a smart sensing

564
00:36:58,040 --> 00:37:04,640
system to improve semiconductor processing and all these types of things.

565
00:37:04,640 --> 00:37:07,120
So that's where the early developments will come.

566
00:37:07,120 --> 00:37:11,320
And there's a real commercial driver for that, which will probably lead to technologies that

567
00:37:11,320 --> 00:37:14,600
then will work in other environments like the home.

568
00:37:14,600 --> 00:37:19,880
I think the other caveat to that is there are, I'm not an expert in this, but I read

569
00:37:19,880 --> 00:37:20,880
a fair bit about it.

570
00:37:20,880 --> 00:37:26,320
And I did read a story about that in South Korea, the amount of robotics use in factories,

571
00:37:26,320 --> 00:37:30,560
etc. is maybe a little bit larger than in some countries, but they're not without their

572
00:37:30,560 --> 00:37:31,560
accidents.

573
00:37:31,560 --> 00:37:37,120
I did hear a story of a robot basically crushing a human worker to death without realizing

574
00:37:37,120 --> 00:37:41,960
the human worker was there, even though it had some sort of limited vision capabilities.

575
00:37:41,960 --> 00:37:45,600
So I'm sure there are a number of barriers beyond the ones we've talked about.

576
00:37:45,600 --> 00:37:47,800
The revolution begins, right?

577
00:37:47,800 --> 00:37:49,480
Well, indeed.

578
00:37:49,480 --> 00:37:52,640
So yeah, it's easy to get excited.

579
00:37:52,640 --> 00:37:54,000
And I think there's stuff to be excited about.

580
00:37:54,000 --> 00:37:56,440
I'm sure there's plenty of barriers to overcome.

581
00:37:56,440 --> 00:38:00,120
But the thing I always think about, how does this tie back to you marketing folks that

582
00:38:00,120 --> 00:38:06,120
are listening is when chat GPT vision became a thing and you could give chat GPT an image

583
00:38:06,120 --> 00:38:11,000
and it knew everything that was in the image and it could help you mark someone's homework

584
00:38:11,000 --> 00:38:13,000
or whatever.

585
00:38:13,000 --> 00:38:18,040
The minute a computer can do that, why does it have to be a static computer that can do

586
00:38:18,040 --> 00:38:19,040
that?

587
00:38:19,040 --> 00:38:24,200
Why can't it be a robot that can now see the world with greater clarity and understanding

588
00:38:24,200 --> 00:38:29,360
than robots could have achieved even maybe three years ago?

589
00:38:29,360 --> 00:38:35,040
The convergence of all these technologies is the types of things that trigger exponential

590
00:38:35,040 --> 00:38:37,080
change, right?

591
00:38:37,080 --> 00:38:43,120
You don't get larger language models without the internet and the development of GPUs and

592
00:38:43,120 --> 00:38:48,200
then the improvements in GPUs, which were themselves driven by the gaming industry trying

593
00:38:48,200 --> 00:38:51,640
to render polygons to make cool looking games.

594
00:38:51,640 --> 00:38:56,520
And these are just a couple of strands that all come together to form this rope that is

595
00:38:56,520 --> 00:39:01,720
this exponential curve that I do believe we're on as it relates to technology improvements

596
00:39:01,720 --> 00:39:06,880
that then you see, well, now that's enabling robotics in this way.

597
00:39:06,880 --> 00:39:09,440
So the software parts are really evolving very quickly.

598
00:39:09,440 --> 00:39:10,440
It's pretty cool.

599
00:39:10,440 --> 00:39:15,760
I'm quite excited about it anyway.

600
00:39:15,760 --> 00:39:17,080
We're going to switch gears slightly now.

601
00:39:17,080 --> 00:39:19,320
We're going to talk rag.

602
00:39:19,320 --> 00:39:23,580
Martin's mentioned it a few times in the episode so far.

603
00:39:23,580 --> 00:39:28,280
And it's also important as it relates even to GPTs because GPTs have a form of rag to

604
00:39:28,280 --> 00:39:32,280
try and help them better answer your questions.

605
00:39:32,280 --> 00:39:33,600
Let's talk a little bit about rag.

606
00:39:33,600 --> 00:39:36,320
We've got a question we've been pondering mine, haven't we?

607
00:39:36,320 --> 00:39:38,640
Where are all the rag chatbots?

608
00:39:38,640 --> 00:39:42,920
So could you start by just letting the listeners know what is rag, why it is important and

609
00:39:42,920 --> 00:39:45,920
why are we thinking about rag chatbots at the moment?

610
00:39:45,920 --> 00:39:46,920
Yeah.

611
00:39:46,920 --> 00:39:50,720
So rag stands for retrieval augmented generation.

612
00:39:50,720 --> 00:39:55,040
And it's a strategy designed to enhance the performance of large language models when

613
00:39:55,040 --> 00:40:01,540
it comes to accessing data that might not be within the training set.

614
00:40:01,540 --> 00:40:08,160
So proprietary data about your company, maybe it's information from your product manuals

615
00:40:08,160 --> 00:40:09,160
or something like that.

616
00:40:09,160 --> 00:40:16,960
And it's a way of limiting hallucinations and feeding the large language model information

617
00:40:16,960 --> 00:40:23,400
that would be maybe outside of the training data because it was more recent than the cutoff

618
00:40:23,400 --> 00:40:24,880
date, something like that.

619
00:40:24,880 --> 00:40:33,240
So there's a lot of hype around this at the moment because lots of companies obviously

620
00:40:33,240 --> 00:40:39,560
want to have chatbots and chat GPT that's connected to your data.

621
00:40:39,560 --> 00:40:41,120
Wouldn't that be wonderful?

622
00:40:41,120 --> 00:40:48,400
But we're just not seeing that roll out into the real world at the speed at which the hype

623
00:40:48,400 --> 00:40:53,600
around LLMs suggests that you would expect to see it.

624
00:40:53,600 --> 00:41:02,840
And there's an interesting example recently where I think we get a bit of an insight into

625
00:41:02,840 --> 00:41:09,040
exactly why we're not seeing these deployed at the pace that you would maybe expect.

626
00:41:09,040 --> 00:41:20,240
And it's actually less to do with rag per se than it is to do with the limitations of

627
00:41:20,240 --> 00:41:22,560
large language models on the whole.

628
00:41:22,560 --> 00:41:29,360
So the great example of Chevrolet, the car manufacturer deployed an AI chatbot and it

629
00:41:29,360 --> 00:41:35,800
was powered by GPT and open AI large language models.

630
00:41:35,800 --> 00:41:40,920
But what became apparent very quickly was despite the fact that it was connected to

631
00:41:40,920 --> 00:41:46,240
the company data and it had this rag integration so it could tell you all about Chevrolet cars

632
00:41:46,240 --> 00:41:52,880
and prices and various models and what have you, it was still susceptible to a technique

633
00:41:52,880 --> 00:41:58,600
of what is effectively hacking the system and it was a technique called prompt injection.

634
00:41:58,600 --> 00:42:03,600
So users were able to manipulate the chatbot into offering ridiculously low prices for

635
00:42:03,600 --> 00:42:09,160
the vehicles, getting it to agree to legally binding offers to sell a car for a dollar.

636
00:42:09,160 --> 00:42:15,720
And this was basically achieved by users in the conversation with the chatbot pretending

637
00:42:15,720 --> 00:42:22,480
to be the manager of a dealership or even the CEO of open AI and instructing the chatbot

638
00:42:22,480 --> 00:42:26,800
to agree to all customer statements or special offers.

639
00:42:26,800 --> 00:42:31,520
And then people were obviously screenshotting these conversations and sharing them online

640
00:42:31,520 --> 00:42:40,760
and pointing out the vulnerabilities of these kind of implementations, I guess is the word

641
00:42:40,760 --> 00:42:42,240
I'm looking for there.

642
00:42:42,240 --> 00:42:47,960
So this just goes to show that why companies aren't rolling these out at speed because

643
00:42:47,960 --> 00:42:52,480
everybody's very excited for them and the promise of personalised experiences and the

644
00:42:52,480 --> 00:42:58,120
promise of having a chat GPT powered by GPT4 that knows everything about your company is

645
00:42:58,120 --> 00:43:04,960
wonderful except they still suffer from the same things that large language models suffer

646
00:43:04,960 --> 00:43:10,640
from and companies do not have the resources available to them to spend so long effectively

647
00:43:10,640 --> 00:43:14,200
red teaming against all of these scenarios.

648
00:43:14,200 --> 00:43:21,680
So basically putting a team of people trying to hack their own bot for almost like, you

649
00:43:21,680 --> 00:43:26,080
know, what do they call it in cyber security where it's like safe hacking.

650
00:43:26,080 --> 00:43:31,680
Yeah, you try and find all the possible holes so you can patch them before your product

651
00:43:31,680 --> 00:43:32,680
goes to market.

652
00:43:32,680 --> 00:43:38,360
But you're right, large language models are so prone to hallucination and error, they're

653
00:43:38,360 --> 00:43:39,540
somewhat unpatchable.

654
00:43:39,540 --> 00:43:42,280
I think that's my interpretation.

655
00:43:42,280 --> 00:43:46,160
And most businesses, I think a lot of businesses have been trying to do this and then they've

656
00:43:46,160 --> 00:43:49,120
realised this is hard.

657
00:43:49,120 --> 00:43:57,240
It's going to be prone to errors and we just commercially can't afford any bad things to

658
00:43:57,240 --> 00:43:58,240
happen.

659
00:43:58,240 --> 00:44:03,720
It's like someone tries to use the chat bot of a particular product and then they follow

660
00:44:03,720 --> 00:44:07,440
the advice the chat bot gives and they get electrocuted because the chat bot forgot to

661
00:44:07,440 --> 00:44:11,760
say something that a human would know to say or made something up that you should never

662
00:44:11,760 --> 00:44:12,760
do, right?

663
00:44:12,760 --> 00:44:16,080
If it was a customer service bot, you could absolutely imagine that.

664
00:44:16,080 --> 00:44:23,340
I think with the company that we were talking about earlier that's come up with this mechanism

665
00:44:23,340 --> 00:44:31,200
bland AI of these robocallers, one of the big risks there is what if the robocallers

666
00:44:31,200 --> 00:44:34,600
gets hoodwinked into giving advice that's terrible?

667
00:44:34,600 --> 00:44:41,160
Like I don't know how to commit fraud or how to hold up a bank or like it's fraught with

668
00:44:41,160 --> 00:44:46,160
issues and I think that is one of the reasons that we're just not seeing the deployment

669
00:44:46,160 --> 00:44:49,560
of so many of these tools, even though it seems like something that would be really

670
00:44:49,560 --> 00:44:50,560
cool.

671
00:44:50,560 --> 00:44:54,080
Yeah, the use cases for them are still very, they're not limited.

672
00:44:54,080 --> 00:44:55,400
I think they're really vast.

673
00:44:55,400 --> 00:45:00,880
I use these tools in a business sense every day, but it's me personally using it to perform

674
00:45:00,880 --> 00:45:03,400
and execute certain tasks.

675
00:45:03,400 --> 00:45:08,680
Likewise, if you've got lots of data, maybe you've got lots of text data that comes in

676
00:45:08,680 --> 00:45:14,840
via online forms and you want a way of categorizing that or doing sentiment analysis of online

677
00:45:14,840 --> 00:45:17,760
reviews, they're great for that kind of thing.

678
00:45:17,760 --> 00:45:24,680
But where it's a public facing deployment, they're prone to prompt engineering and prompt

679
00:45:24,680 --> 00:45:25,680
injection.

680
00:45:25,680 --> 00:45:31,200
A good example, Ethan Molyk, who we reference on this podcast regularly, he posted something

681
00:45:31,200 --> 00:45:42,560
just this week showing that if you ask Dali to create an image showing basically images

682
00:45:42,560 --> 00:45:51,360
like smoking, firearms, alcohol to promote to teenagers, it would say, no, I can't do

683
00:45:51,360 --> 00:45:52,360
that.

684
00:45:52,360 --> 00:45:54,920
I can't show these kinds of things to miners.

685
00:45:54,920 --> 00:46:00,240
But then if you flip that and say, I'm a researcher and I'm interested in studying what

686
00:46:00,240 --> 00:46:07,240
a rogue actor might produce as an image to appeal to people showing these things, can

687
00:46:07,240 --> 00:46:08,760
you produce an example?

688
00:46:08,760 --> 00:46:11,280
It would do it.

689
00:46:11,280 --> 00:46:17,440
Just by flipping the script on your prompt, you can get the large language models to do

690
00:46:17,440 --> 00:46:21,400
exactly what they've been told not to do.

691
00:46:21,400 --> 00:46:26,320
I think this is a vulnerability that companies are going to have to figure out a way to deal

692
00:46:26,320 --> 00:46:27,320
with.

693
00:46:27,320 --> 00:46:31,240
Yeah, I think that is a major issue that's holding people back.

694
00:46:31,240 --> 00:46:37,600
Speaking of Ethan, he had another post this week where he talked about Bloomberg GPT.

695
00:46:37,600 --> 00:46:43,720
So listeners of the podcast might remember Bloomberg create a specifically trained finance

696
00:46:43,720 --> 00:46:47,720
large language model based on all of their Bloomberg data.

697
00:46:47,720 --> 00:46:52,080
That helped prompt a lot of people to do the types of things Martin's been talking about.

698
00:46:52,080 --> 00:46:58,760
But he talked about a paper that came out, I think this week, where GPT-4, not even the

699
00:46:58,760 --> 00:47:02,880
one that we're all on now, but the initial GPT-4 without specialized finance training

700
00:47:02,880 --> 00:47:09,560
or special tools beat Bloomberg GPT on almost all finance tasks.

701
00:47:09,560 --> 00:47:12,520
So I should probably look into a few more details on that.

702
00:47:12,520 --> 00:47:17,160
I would guess that probably the underlying model that Bloomberg GPT was built on was

703
00:47:17,160 --> 00:47:19,640
maybe GPT 3.5, for example.

704
00:47:19,640 --> 00:47:26,760
Yeah, I believe it was because you can't fine tune GPT-4.

705
00:47:26,760 --> 00:47:27,760
Right.

706
00:47:27,760 --> 00:47:29,680
So that is really interesting.

707
00:47:29,680 --> 00:47:33,960
It's not the first time GPT-4 has been able to be a specialized model that was trained

708
00:47:33,960 --> 00:47:34,960
on specialized data.

709
00:47:34,960 --> 00:47:39,320
I mean, goodness knows what was in the GPT-4 training data set, like every bit of information

710
00:47:39,320 --> 00:47:41,840
that humanity's ever had.

711
00:47:41,840 --> 00:47:48,320
But the fact that there's an ongoing discussion in this world about what's going to be better,

712
00:47:48,320 --> 00:47:51,960
massively general models that are trained on huge amounts of information and can basically

713
00:47:51,960 --> 00:47:57,760
do everything, just use GPT-4 for everything, for example, or specialized trained models

714
00:47:57,760 --> 00:48:02,200
that are really good in a very, very specific domain that you should use for specific use

715
00:48:02,200 --> 00:48:10,160
cases, think a doctor GPT, and that would be better than a general model.

716
00:48:10,160 --> 00:48:14,400
But GPT-4 seems to do surprisingly well when it's put to the test against a lot of these

717
00:48:14,400 --> 00:48:19,800
fine-tuned models, which again begs the question, why the only reason I can think of then to

718
00:48:19,800 --> 00:48:27,600
create a rack-driven chatbot is if there's information you know for sure is not in the

719
00:48:27,600 --> 00:48:31,040
training data, as you said, Martin, like information on your own products.

720
00:48:31,040 --> 00:48:33,120
This thing on my thing is broken.

721
00:48:33,120 --> 00:48:34,720
What is the best way to fix that thing?

722
00:48:34,720 --> 00:48:37,160
And it's like, well, of course it probably doesn't know.

723
00:48:37,160 --> 00:48:42,880
But yeah, I think it's a great question that you asked because it's such a cool and interesting

724
00:48:42,880 --> 00:48:46,400
use case, but we just haven't really seen anybody doing it.

725
00:48:46,400 --> 00:48:51,400
And I think we talk as well about assistants.

726
00:48:51,400 --> 00:48:55,960
Why is Google Assistant still so kind of rubbish?

727
00:48:55,960 --> 00:49:00,920
Why is Alexa still stuck in what feels like 1962?

728
00:49:00,920 --> 00:49:02,680
Siri the same.

729
00:49:02,680 --> 00:49:05,600
And the people at these companies are super smart.

730
00:49:05,600 --> 00:49:07,280
They see what's going on.

731
00:49:07,280 --> 00:49:12,000
Why don't we just get Google Assistant upgraded large language model style?

732
00:49:12,000 --> 00:49:13,000
Why don't we have it now?

733
00:49:13,000 --> 00:49:15,040
Like, why didn't we have it six months ago?

734
00:49:15,040 --> 00:49:19,800
And I think one of the reasons is the things you're talking about, the red teaming.

735
00:49:19,800 --> 00:49:26,760
You can't have Google Assistant going off the deep end, giving advice that it shouldn't

736
00:49:26,760 --> 00:49:34,040
or cascading social biases or being completely tripped up by prompt injection.

737
00:49:34,040 --> 00:49:39,680
Imagine the press when people record snippets of a conversation with Alexa where they managed

738
00:49:39,680 --> 00:49:43,640
to get it to say something highly offensive about a specific group of people and what

739
00:49:43,640 --> 00:49:47,120
that would do to Amazon's brand.

740
00:49:47,120 --> 00:49:53,240
So the same things that are plaguing these chatbots are, I think, also holding back other

741
00:49:53,240 --> 00:49:58,760
implementations of what would be really powerful tools like better assistants because of all

742
00:49:58,760 --> 00:50:02,800
the edge cases of how easy it is to fool them to do stuff they shouldn't or because sometimes

743
00:50:02,800 --> 00:50:04,960
they just make stuff up and talk rubbish.

744
00:50:04,960 --> 00:50:05,960
Right.

745
00:50:05,960 --> 00:50:10,200
So we've just got a couple of stories now as we get into the last 10 minutes or so.

746
00:50:10,200 --> 00:50:14,400
One thing that caught our eye this week was the perplexity funding announcement.

747
00:50:14,400 --> 00:50:20,640
So perplexity is an AI company that's sort of somewhere between feels like to me, Martin's

748
00:50:20,640 --> 00:50:23,800
going to tell us more in a minute because I'm not an expert, but it feels like a mixture

749
00:50:23,800 --> 00:50:26,280
of Google and chat GPT all smashed together.

750
00:50:26,280 --> 00:50:32,040
And they just got $73.6 million, which puts their valuation at just over half a billion,

751
00:50:32,040 --> 00:50:34,440
which is not insignificant.

752
00:50:34,440 --> 00:50:38,320
I've been trying to use perplexity and not been able to get myself fully into it, although

753
00:50:38,320 --> 00:50:42,760
I've done better this week because Martin gave me some coaching, but you mind, I think

754
00:50:42,760 --> 00:50:45,360
you think perplexity is going to have a big impact.

755
00:50:45,360 --> 00:50:50,160
So can you tell us in the listeners, like what it is, what you use it for and why you

756
00:50:50,160 --> 00:50:51,640
think it's so awesome?

757
00:50:51,640 --> 00:50:52,640
Yeah.

758
00:50:52,640 --> 00:50:54,760
So it's a search engine, right?

759
00:50:54,760 --> 00:51:01,600
First and foremost, it's a search engine and you ask it questions like you would do Google

760
00:51:01,600 --> 00:51:03,360
or anything else.

761
00:51:03,360 --> 00:51:07,840
But the responses that it gives you rather than just giving you a list of places to then

762
00:51:07,840 --> 00:51:10,960
go off to and find information.

763
00:51:10,960 --> 00:51:11,960
It doesn't do that.

764
00:51:11,960 --> 00:51:15,080
It presents the information like chat GPT does.

765
00:51:15,080 --> 00:51:17,300
It gives you a written answer.

766
00:51:17,300 --> 00:51:21,120
Now it does also give you links to the places that you want to go.

767
00:51:21,120 --> 00:51:25,520
So if you want to use it in the same way that you would do a navigational search on Google,

768
00:51:25,520 --> 00:51:26,520
right?

769
00:51:26,520 --> 00:51:31,440
You've got Google Chrome open, the search bar at the top of the address bar is defaulted

770
00:51:31,440 --> 00:51:32,440
to Google search.

771
00:51:32,440 --> 00:51:38,440
If I type in biostrata, I don't put your website in, it's going to bring up a Google listing

772
00:51:38,440 --> 00:51:42,800
page and presumably biostrata website at the top.

773
00:51:42,800 --> 00:51:46,240
I click on that and go to your website.

774
00:51:46,240 --> 00:51:47,400
Perplexity will do that.

775
00:51:47,400 --> 00:51:52,560
I go on there, I type in biostrata.

776
00:51:52,560 --> 00:51:54,240
I do my search.

777
00:51:54,240 --> 00:51:57,120
There we see I can, I'm doing it right now.

778
00:51:57,120 --> 00:51:58,440
Biostrata is linked there.

779
00:51:58,440 --> 00:52:03,360
I've got a link to your LinkedIn page and beneath that is a block of text that tells

780
00:52:03,360 --> 00:52:06,600
me all about biostrata in great detail.

781
00:52:06,600 --> 00:52:13,200
It's giving me a summary of the organisation, what you're known for, the industry that you're

782
00:52:13,200 --> 00:52:16,600
in and what have you.

783
00:52:16,600 --> 00:52:19,520
Originally I was using this not for navigational searches.

784
00:52:19,520 --> 00:52:23,840
I was using it for, I wanted, I was researching a topic.

785
00:52:23,840 --> 00:52:29,000
So I would do a, I would ask a more complex question, a long tail keyword effectively.

786
00:52:29,000 --> 00:52:33,520
And I would get it to maybe explain a new story that had just occurred or tell me about

787
00:52:33,520 --> 00:52:36,120
the key features of this product demonstration.

788
00:52:36,120 --> 00:52:42,480
But increasingly I'm using it, well, in fact, last week I made the switch to perplexity

789
00:52:42,480 --> 00:52:47,400
being my primary search engine across devices.

790
00:52:47,400 --> 00:52:53,600
It's my default search engine on my browsers and I use it for everything now.

791
00:52:53,600 --> 00:52:55,640
And use cases for it are really varied.

792
00:52:55,640 --> 00:53:04,520
So the other day it was FA Cup round, round three and FA Cup games are broadcast across

793
00:53:04,520 --> 00:53:06,720
the BBC on ITV.

794
00:53:06,720 --> 00:53:10,880
I can never remember whether they're on TNT Sports or Sky Sports.

795
00:53:10,880 --> 00:53:12,200
I just don't know.

796
00:53:12,200 --> 00:53:15,820
And if you ever do that search on Google, it's infuriating, right?

797
00:53:15,820 --> 00:53:22,540
So if you do the search on Google and say, what channel is Manchester United versus Wigan

798
00:53:22,540 --> 00:53:30,400
on Monday, the listings will bring up basically local newspapers, The Sun, The Daily Mirror.

799
00:53:30,400 --> 00:53:36,880
And if you click into the story, you then get a newspaper advert or newspaper page filled

800
00:53:36,880 --> 00:53:37,880
with adverts.

801
00:53:37,880 --> 00:53:44,120
And the last sentence on the story tells you that it's on ITV at 7pm.

802
00:53:44,120 --> 00:53:47,240
And you go, okay, that's the information I wanted.

803
00:53:47,240 --> 00:53:49,240
With perplexity, I say, what channel is it on?

804
00:53:49,240 --> 00:53:50,240
And it tells me.

805
00:53:50,240 --> 00:53:51,240
Great.

806
00:53:51,240 --> 00:53:52,240
Good start.

807
00:53:52,240 --> 00:53:58,480
I had a use case earlier this week where it was a sales call.

808
00:53:58,480 --> 00:54:04,200
So someone had booked a meeting with me, a new prospect, potential new client.

809
00:54:04,200 --> 00:54:06,080
I did a Google search.

810
00:54:06,080 --> 00:54:07,400
Well, I did a Google search.

811
00:54:07,400 --> 00:54:10,000
You see, it's so ingrained in me.

812
00:54:10,000 --> 00:54:17,920
I did a search on perplexity and just typed in the company name and put their location.

813
00:54:17,920 --> 00:54:20,920
And it brings up the website much like searching for you.

814
00:54:20,920 --> 00:54:22,760
And then it gives me a bit of a blurb.

815
00:54:22,760 --> 00:54:26,200
So immediately I understand who this company is, what they do.

816
00:54:26,200 --> 00:54:27,760
I haven't had to go to their about page.

817
00:54:27,760 --> 00:54:31,520
I'm like straight in, who are these people?

818
00:54:31,520 --> 00:54:32,840
And it has follow-up chat.

819
00:54:32,840 --> 00:54:39,880
So in the same way that you have chat GPT conversations, you can ask follow-up questions.

820
00:54:39,880 --> 00:54:46,620
Much like the search generative experience in Google has, perplexity has this baked in.

821
00:54:46,620 --> 00:54:50,040
You can upload images and it can examine images.

822
00:54:50,040 --> 00:54:51,920
It does image search as well.

823
00:54:51,920 --> 00:54:57,280
So if you're researching a holiday, you can ask it to put together a travel itinerary

824
00:54:57,280 --> 00:54:58,280
for you.

825
00:54:58,280 --> 00:55:00,180
Like you would ask chat GPT.

826
00:55:00,180 --> 00:55:01,260
It will do that.

827
00:55:01,260 --> 00:55:04,240
And then the whole user experience of it is really nice.

828
00:55:04,240 --> 00:55:09,680
It's hard to articulate fully, but when you use it, it's just very pleasant.

829
00:55:09,680 --> 00:55:10,680
Do you know what?

830
00:55:10,680 --> 00:55:11,680
It's clean.

831
00:55:11,680 --> 00:55:12,680
That's what I like about it.

832
00:55:12,680 --> 00:55:18,080
You make a really interesting point about the go to like tabloid newspaper website and

833
00:55:18,080 --> 00:55:22,080
it's full of images and I can't find the information I want until I've scrolled through the images,

834
00:55:22,080 --> 00:55:24,760
because of course that's how they get paid.

835
00:55:24,760 --> 00:55:29,940
And even Google search is actually kind of a bit cluttered with the ads and shopping

836
00:55:29,940 --> 00:55:34,640
ads and you've got your year 10 links, but you've got a little bit of copy under them.

837
00:55:34,640 --> 00:55:39,600
There's something kind of nice about just reading a little couple of paragraphs of text

838
00:55:39,600 --> 00:55:43,760
created just for you where all the links are like references and assigned to the paper.

839
00:55:43,760 --> 00:55:46,960
Like there's little button you can click on if you want to see the link, but it's just

840
00:55:46,960 --> 00:55:47,960
clean.

841
00:55:47,960 --> 00:55:48,960
I like that about it.

842
00:55:48,960 --> 00:55:49,960
Yeah.

843
00:55:49,960 --> 00:55:52,520
And it has integration with GPT-4.

844
00:55:52,520 --> 00:55:59,760
So if you have perplexity pro, I think it's called perplexity pro $200 a year or $20 a

845
00:55:59,760 --> 00:56:06,420
month, you get GPT-4, you get Claude 2, they've integrated Gemini.

846
00:56:06,420 --> 00:56:08,420
You can use it as a writing tool.

847
00:56:08,420 --> 00:56:11,000
So you can just have it rather than as a search engine.

848
00:56:11,000 --> 00:56:17,840
It's got a feature where you can just go into writing mode and just use it as a chat much

849
00:56:17,840 --> 00:56:20,920
like the others and you can choose your model as well.

850
00:56:20,920 --> 00:56:27,840
So yeah, I'm completely into it and the level of funding that it's got doesn't surprise

851
00:56:27,840 --> 00:56:29,120
me.

852
00:56:29,120 --> 00:56:35,680
There were some big players in tech posting this week and in the weeks prior to the announcement

853
00:56:35,680 --> 00:56:43,600
that they'd switched to being or switched it to having as their primary search engine.

854
00:56:43,600 --> 00:56:46,400
Jeff Bezos has put a load of money behind it.

855
00:56:46,400 --> 00:56:52,920
Obviously he's got an interesting backing horse against Google in this domain.

856
00:56:52,920 --> 00:56:55,080
But yeah, I'm fully into it.

857
00:56:55,080 --> 00:56:56,080
I think it's great.

858
00:56:56,080 --> 00:56:57,080
Yeah.

859
00:56:57,080 --> 00:57:01,400
So as you said, there's a paid version which adds maybe some additional complexity and

860
00:57:01,400 --> 00:57:02,400
remove some limits.

861
00:57:02,400 --> 00:57:05,200
But if you want to try perplexity, you can go try it now because it's free.

862
00:57:05,200 --> 00:57:07,360
You have to create a login, but there's a free version, right?

863
00:57:07,360 --> 00:57:08,360
Yes.

864
00:57:08,360 --> 00:57:09,360
Yeah, there is.

865
00:57:09,360 --> 00:57:10,360
And it gives you a lot of capability just with that.

866
00:57:10,360 --> 00:57:11,360
Yeah.

867
00:57:11,360 --> 00:57:14,800
I checked by Strateroom while you were talking and I was quite interested and impressed with

868
00:57:14,800 --> 00:57:22,840
its description because it didn't just pull some obvious text off of like our homepage.

869
00:57:22,840 --> 00:57:27,480
It pulled information across multiple pages of our site and also some information about

870
00:57:27,480 --> 00:57:32,720
by Strateroom that you'd find on other sites, which I think is really interesting.

871
00:57:32,720 --> 00:57:37,120
I talk a lot in the workshops that I do about different use cases that I like and because

872
00:57:37,120 --> 00:57:41,360
I have a business development role at by Strateroom, among other things, I do due diligence on

873
00:57:41,360 --> 00:57:47,040
leads where usually I'll say, tell me about company X and I'll ask Claude, chat GPT,

874
00:57:47,040 --> 00:57:48,040
Bard.

875
00:57:48,040 --> 00:57:51,920
But now perplexity is going to go into that list because I'm willing to guess that it

876
00:57:51,920 --> 00:57:55,520
will probably do a better job than all of them, to be honest.

877
00:57:55,520 --> 00:57:57,360
So I'm going to slot it in there.

878
00:57:57,360 --> 00:57:58,360
Cool.

879
00:57:58,360 --> 00:57:59,360
So there you are, listeners.

880
00:57:59,360 --> 00:58:04,920
Go and have a play with perplexity as a research tool as a potential replacement for search.

881
00:58:04,920 --> 00:58:06,720
It's worth checking out.

882
00:58:06,720 --> 00:58:07,720
Right.

883
00:58:07,720 --> 00:58:10,360
We want to respect our listeners' time as always.

884
00:58:10,360 --> 00:58:14,640
So we've got, I think the most important story for us to focus on is the CES announcements

885
00:58:14,640 --> 00:58:16,200
because there's a bit of fun in there, Martin.

886
00:58:16,200 --> 00:58:19,200
So I'm going to whip the listeners through a couple of stories we won't have time to

887
00:58:19,200 --> 00:58:22,040
go into detail on, but hopefully it's still interesting.

888
00:58:22,040 --> 00:58:26,840
So people who are relying on us for updates in this space should know that the New York

889
00:58:26,840 --> 00:58:34,040
Times sued OpenAI for copyright infringement by using its content to train the models.

890
00:58:34,040 --> 00:58:37,920
It's proving to be quite an interesting and dynamic and complex case that will probably

891
00:58:37,920 --> 00:58:40,200
take a number of months to resolve.

892
00:58:40,200 --> 00:58:44,480
They're not legal experts, but it's worth paying attention to this case because it will

893
00:58:44,480 --> 00:58:50,920
define the future of how these models get trained and also how publishing houses like

894
00:58:50,920 --> 00:58:58,400
the New York Times are compensated for the content that they're producing to fuel large

895
00:58:58,400 --> 00:58:59,680
language models.

896
00:58:59,680 --> 00:59:03,000
So it's kind of an interesting one to pay attention to.

897
00:59:03,000 --> 00:59:09,040
Another quick story was that we've talked earlier about how Bing is now being rebranded

898
00:59:09,040 --> 00:59:15,080
as Microsoft Copilot for a lot of use cases, and there's now a Copilot app on iOS and

899
00:59:15,080 --> 00:59:16,720
Android, which is quite fun to use.

900
00:59:16,720 --> 00:59:20,640
It's kind of a bit chat GPT-like, worth going and having a play with.

901
00:59:20,640 --> 00:59:24,280
And as part of that, Microsoft is going to be making the first major change to the Windows

902
00:59:24,280 --> 00:59:29,920
keyboard in 30 years by adding a new Copilot button, which is kind of interesting.

903
00:59:29,920 --> 00:59:34,440
And then the last thing that we saw was that Apple released a research paper called LLM

904
00:59:34,440 --> 00:59:39,280
in a flash, which is basically running a large language model directly on a smartphone with

905
00:59:39,280 --> 00:59:41,920
smartphone hardware.

906
00:59:41,920 --> 00:59:47,000
And for those of you that are very privacy conscious, all of the large language models

907
00:59:47,000 --> 00:59:49,240
that most of us are using are cloud-based.

908
00:59:49,240 --> 00:59:55,160
You put something into chat GPT and off it goes to chat GPT and open AI servers, where

909
00:59:55,160 --> 01:00:00,040
in theory they could always look at the information that you're sharing with them.

910
01:00:00,040 --> 01:00:04,280
But there is a movement to get large language models running on device.

911
01:00:04,280 --> 01:00:07,960
So you have a large language model installed on your computer and it runs on your computer

912
01:00:07,960 --> 01:00:10,160
and doesn't send any information anywhere.

913
01:00:10,160 --> 01:00:14,480
And what Apple are doing is looking at how they can actually even make that work on a

914
01:00:14,480 --> 01:00:15,680
smartphone.

915
01:00:15,680 --> 01:00:18,360
So that's pretty cool and interesting as well.

916
01:00:18,360 --> 01:00:23,480
With that, let's get into our very last story then, which is some of the unique AI-powered

917
01:00:23,480 --> 01:00:26,480
products that we saw at CES this year.

918
01:00:26,480 --> 01:00:29,040
Take us through some of the cool things you saw, Mike.

919
01:00:29,040 --> 01:00:33,100
The first one that really caught my eye is one called Wisp, which is an assistive technology

920
01:00:33,100 --> 01:00:35,160
from a Netherlands-based startup.

921
01:00:35,160 --> 01:00:40,360
And what this does is it converts whispered speech and any kind of affected speech into

922
01:00:40,360 --> 01:00:42,680
clear natural voice.

923
01:00:42,680 --> 01:00:49,080
And it's designed to help people that have severe stutter, throat cancer, vocal cord

924
01:00:49,080 --> 01:00:54,560
paralysis, or anything else that might affect natural speech patterns.

925
01:00:54,560 --> 01:01:00,360
The applications AI technology is language independent and works on both iOS and Android.

926
01:01:00,360 --> 01:01:02,320
So I think that's a really cool one.

927
01:01:02,320 --> 01:01:03,320
Yeah, I love that.

928
01:01:03,320 --> 01:01:07,960
What a good and productive implementation of AI.

929
01:01:07,960 --> 01:01:08,960
Yeah.

930
01:01:08,960 --> 01:01:18,080
And staying on the assistive AIs, the Starkey Genesis AI developed by Starkey Labs, they

931
01:01:18,080 --> 01:01:23,680
created this Genesis AI, which is a hearing aid technology that includes an onboard deep

932
01:01:23,680 --> 01:01:25,640
neural network accelerator engine.

933
01:01:25,640 --> 01:01:32,560
It promises wear as the ability to hear soft sounds without noise, more natural distinguishing

934
01:01:32,560 --> 01:01:38,560
of words and speech, and generally makes it less work for you to listen.

935
01:01:38,560 --> 01:01:46,560
So the device's receiver-in-canal hearing aid lasts up to 51 hours and is waterproof.

936
01:01:46,560 --> 01:01:51,320
So better hearing aids thanks to neural networks being able to figure out the difference between

937
01:01:51,320 --> 01:01:54,040
the things you want to hear and all the background noise.

938
01:01:54,040 --> 01:01:55,960
Again, very welcome.

939
01:01:55,960 --> 01:01:57,360
What's not to love?

940
01:01:57,360 --> 01:02:04,100
Volkswagen have announced plans to add an AI powered chat bot into all Volkswagen models

941
01:02:04,100 --> 01:02:08,540
equipped with its IDA voice assistant.

942
01:02:08,540 --> 01:02:15,520
So this is based on the software company Serence Chat Pro, which I'm not familiar with, haven't

943
01:02:15,520 --> 01:02:21,640
heard of them, and works with OpenAI's foundational models as well.

944
01:02:21,640 --> 01:02:26,480
And it's going to roll out across Europe starting in the second quarter.

945
01:02:26,480 --> 01:02:31,320
So nice little upgrade for Volkswagen owners there.

946
01:02:31,320 --> 01:02:32,320
About time.

947
01:02:32,320 --> 01:02:33,320
Knight Rider's coming.

948
01:02:33,320 --> 01:02:39,160
I need to be able to talk with my car, A, for fun, but B, as we've talked about previously,

949
01:02:39,160 --> 01:02:43,840
how awesome to be able to drive back from a meeting, have a conversation with the tool

950
01:02:43,840 --> 01:02:48,240
about the meeting that you had and in effect, have it ask you great questions about how

951
01:02:48,240 --> 01:02:51,880
the meeting went and then take all the notes for you as your thoughts come out of your

952
01:02:51,880 --> 01:02:53,360
mind after your meeting.

953
01:02:53,360 --> 01:02:55,800
Very excited for some of those use cases.

954
01:02:55,800 --> 01:02:58,680
Yeah, that is one that I think is great.

955
01:02:58,680 --> 01:03:03,400
And finally, the most important, I think we can all agree that this is the one that everybody's

956
01:03:03,400 --> 01:03:04,640
been waiting for.

957
01:03:04,640 --> 01:03:09,520
This came from Gluckskind Rosa.

958
01:03:09,520 --> 01:03:16,480
I don't know if I've said that correctly or not, but they've created what's known as the

959
01:03:16,480 --> 01:03:18,960
Fractional Nanny.

960
01:03:18,960 --> 01:03:24,720
It's an AI powered pram designed to make parenting effortless and enjoyable with cutting edge

961
01:03:24,720 --> 01:03:27,360
robotics and intelligent technology.

962
01:03:27,360 --> 01:03:32,960
So the pram or the stroller, as they would say in the US, features automatic braking

963
01:03:32,960 --> 01:03:37,080
and soothing white noise in a lightweight design.

964
01:03:37,080 --> 01:03:42,600
There had to be a little bit of an outside use case there.

965
01:03:42,600 --> 01:03:43,960
Do you know my favorite thing about that?

966
01:03:43,960 --> 01:03:50,200
As you lifting from their own materials quite clearly, an AI powered pram that would make

967
01:03:50,200 --> 01:03:55,680
parenting effortless and enjoyable, I think it's going to take a little bit more than

968
01:03:55,680 --> 01:04:00,400
a bit of white noise and some automatic brakes on a pram to make parenting effortless, but

969
01:04:00,400 --> 01:04:03,400
I appreciate the sentiment.

970
01:04:03,400 --> 01:04:07,320
Yeah, the marketing department were a little bit overboard.

971
01:04:07,320 --> 01:04:08,320
Yeah.

972
01:04:08,320 --> 01:04:09,840
Come on, we've got the features.

973
01:04:09,840 --> 01:04:11,120
What are the benefits of this?

974
01:04:11,120 --> 01:04:12,920
Well, it breaks on its own, doesn't it?

975
01:04:12,920 --> 01:04:13,920
That's a good thing.

976
01:04:13,920 --> 01:04:15,480
No, we've got to think deeper than that.

977
01:04:15,480 --> 01:04:17,680
It makes parenting effortless.

978
01:04:17,680 --> 01:04:18,840
Do you think we went too far?

979
01:04:18,840 --> 01:04:20,520
No, get it in the press release.

980
01:04:20,520 --> 01:04:21,520
Well.

981
01:04:21,520 --> 01:04:22,520
Why weren't they asleep?

982
01:04:22,520 --> 01:04:23,520
Why am I so sleep deprived?

983
01:04:23,520 --> 01:04:25,520
Why is this awful?

984
01:04:25,520 --> 01:04:26,520
So there you have it.

985
01:04:26,520 --> 01:04:33,380
Well, at least on the artificially intelligent marketing podcast, we can say with some confidence

986
01:04:33,380 --> 01:04:37,920
that we make it effortless for you, dear listeners, to stay up to date on all the things that

987
01:04:37,920 --> 01:04:43,280
are going on in the world of AI, both cool, interesting stuff like gluk-skinned Rosa and

988
01:04:43,280 --> 01:04:47,200
robotics and hopefully a bunch of stuff about actual tools that you can use in marketing

989
01:04:47,200 --> 01:04:50,280
to make your life easier, et cetera.

990
01:04:50,280 --> 01:04:53,140
If you do enjoy the podcast, share it with a friend.

991
01:04:53,140 --> 01:04:54,560
Maybe they'd also like to enjoy.

992
01:04:54,560 --> 01:04:57,040
Please share our stuff on social as well if you like it.

993
01:04:57,040 --> 01:05:00,320
All helps to get the word out and get more people learning about the cool things they

994
01:05:00,320 --> 01:05:02,200
could do with AI.

995
01:05:02,200 --> 01:05:06,040
Other than that, Martin, I shall look forward to speaking to you again soon for our next

996
01:05:06,040 --> 01:05:07,040
episode.

997
01:05:07,040 --> 01:05:08,720
Looking forward to it already.

998
01:05:08,720 --> 01:05:09,720
Cheers, mate.

999
01:05:09,720 --> 01:05:10,720
Bye.

1000
01:05:10,720 --> 01:05:14,460
Thank you for listening to artificially intelligent marketing.

1001
01:05:14,460 --> 01:05:20,520
To stay on top of the latest trends, tips and tools in the world of marketing AI, be

1002
01:05:20,520 --> 01:05:22,280
sure to subscribe.

1003
01:05:22,280 --> 01:05:39,720
We look forward to seeing you again next week.