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 --> 00:00:22,480
results from your marketing efforts. Now let's join our hosts, Paul Avery and Martin Broadhurst.

4
00:00:22,480 --> 00:00:29,340
Hello everyone. Welcome to episode 45 of Artificially Intelligent Marketing. It's Paul Avery here

5
00:00:29,340 --> 00:00:34,660
joined as always by the fantabulous Martin Broadhurst. Martin, how are you sir?

6
00:00:34,660 --> 00:00:42,040
Ready and raring to go. All good here in sunny Derby. Weather's good and spending my, what

7
00:00:42,040 --> 00:00:47,500
is it, Sunday afternoon talking AI with you. What better way to spend a weekend is there

8
00:00:47,500 --> 00:00:48,500
than that?

9
00:00:48,500 --> 00:00:52,160
Says, yeah, I prefer not to respond to that question, honestly, because I think it's probably

10
00:00:52,160 --> 00:00:56,300
quite a lot of things. No, that's not true, dude. Listen, we are here as always to give

11
00:00:56,300 --> 00:01:00,760
you the lowdown of what's been going on in the world of AI so that you can leverage all

12
00:01:00,760 --> 00:01:05,400
the stuff we spent hours reading about in your marketing efforts without having to do

13
00:01:05,400 --> 00:01:09,520
all that reading yourselves. So that's the plan. We're going to get straight into it

14
00:01:09,520 --> 00:01:12,600
because we've got quite a few interesting stories and stuff to cover today. I think

15
00:01:12,600 --> 00:01:16,720
probably the first thing to talk about Martin is some of the experiments you've been running

16
00:01:16,720 --> 00:01:26,160
this week using AI for data analysis, potentially not quite cracked up quite what it's all cracked

17
00:01:26,160 --> 00:01:31,500
up to be, maybe not as useful as you've been hoping it would be in its current form. What's

18
00:01:31,500 --> 00:01:32,920
been going on this week?

19
00:01:32,920 --> 00:01:39,680
I'll be flying out to Phoenix, Arizona for the marketing analytics summit and I'm presenting

20
00:01:39,680 --> 00:01:45,200
on generative AI tools for analysts. I've been trying out lots of different tools and

21
00:01:45,200 --> 00:01:51,880
yeah, it's been a fascinating set of experiments because even though we've had an update from

22
00:01:51,880 --> 00:01:58,960
OpenAI this week or was it last week or very recently where they said data analytics is

23
00:01:58,960 --> 00:02:03,920
improved, I've got some new features, you can edit tabular data, the charts are better,

24
00:02:03,920 --> 00:02:10,260
all of that kind of stuff has improved. There is still, from my reading and through these

25
00:02:10,260 --> 00:02:16,360
experiments I've been doing, a fundamental flaw in extracting insights from these tools.

26
00:02:16,360 --> 00:02:25,840
And a really clear example of that was I uploaded a CSV file, it had 130,000 rows, each row

27
00:02:25,840 --> 00:02:34,000
was an entry from a customer satisfaction survey and I asked it to run a statistical

28
00:02:34,000 --> 00:02:41,320
analysis and just kind of look at the data and tell me about it. And it does a huge amount

29
00:02:41,320 --> 00:02:46,480
and it takes about 90 seconds for it to go through it. And the approach is really fascinating

30
00:02:46,480 --> 00:02:52,040
because what it does is it starts off by describing the data, so it kind of understands the data

31
00:02:52,040 --> 00:02:57,860
set that you've got and what you're working with and then it lays out a analysis plan.

32
00:02:57,860 --> 00:03:02,520
So here's what I think we can do with that data and here's how I'll go about doing it

33
00:03:02,520 --> 00:03:07,720
and then it goes about doing it. So it gets straight into it, makes all of these charts

34
00:03:07,720 --> 00:03:15,120
and these visualisations and then starts giving you insights. The charts and the visualisations

35
00:03:15,120 --> 00:03:26,000
so far as I've seen are really good and accurate but it does fall down on the insights piece.

36
00:03:26,000 --> 00:03:33,000
One very simple example of this was a bar chart where it had grouped together these

37
00:03:33,000 --> 00:03:39,040
passengers by age. So it's created a bunch of age cohorts first and then it's run the

38
00:03:39,040 --> 00:03:46,240
analysis on the cohort data and each of these bar charts had two bars or each of the cohorts

39
00:03:46,240 --> 00:03:50,280
had two bars and it was the level of satisfaction. So were they satisfied or dissatisfied and

40
00:03:50,280 --> 00:03:56,200
it was a simple count of those levels. And on the first one for the first age cohort

41
00:03:56,200 --> 00:04:03,720
the bars were totally uneven. Satisfied was like 5000 and dissatisfied is about 1000.

42
00:04:03,720 --> 00:04:11,240
So it's a massive difference in levels. And the description it gave is in the 0 to 18

43
00:04:11,240 --> 00:04:18,800
cohort you can see that things are quite balanced. Very evidently they're not balanced at all

44
00:04:18,800 --> 00:04:23,880
but it gets worse because on the next one, the next cohort, it just gets the figures

45
00:04:23,880 --> 00:04:30,920
completely wrong and it says the satisfied are higher than the dissatisfied where actually

46
00:04:30,920 --> 00:04:36,600
it's vice versa. So the data that you're looking at is accurate in terms of the tables and

47
00:04:36,600 --> 00:04:46,140
the charts but the insight gleaned from it is just completely wrong. And this extends

48
00:04:46,140 --> 00:04:54,000
to other tools as well. So if you have a look at Copilot for Excel which has been rolled

49
00:04:54,000 --> 00:05:01,760
out across, well Copilot's been rolled out across the 365, Office 365 suite now. In Copilot

50
00:05:01,760 --> 00:05:08,720
for Excel you have to make your data a table and then you can start interrogating it with

51
00:05:08,720 --> 00:05:14,160
natural language. First thing to note, if your table has more than two million cells

52
00:05:14,160 --> 00:05:20,320
of data, Copilot can't work with it. So straight away it will say this table is too big, seems

53
00:05:20,320 --> 00:05:28,640
like a bit of a flaw. But then I tried it with a spreadsheet containing 40 rows of data

54
00:05:28,640 --> 00:05:36,880
of employee data. This was all dummy data right. But I asked it a very simple question.

55
00:05:36,880 --> 00:05:40,560
I asked it how many male employees did we have and there was a column that just had

56
00:05:40,560 --> 00:05:44,640
people labelled as male, female. That was the data that we were working with in this

57
00:05:44,640 --> 00:05:52,360
dummy data set. Now the first column, column A1, well cell A1 says employee ID. So it's

58
00:05:52,360 --> 00:05:56,800
very obvious, like the type of data that we're doing. I didn't feel like it needed to be

59
00:05:56,800 --> 00:06:02,720
described to the model. I asked it to tell me how many male employees there were and

60
00:06:02,720 --> 00:06:07,120
it says well I couldn't tell you that because I don't have access to that information. Maybe

61
00:06:07,120 --> 00:06:11,880
you can put me in the right direction. And if you then say well if you look in this column

62
00:06:11,880 --> 00:06:19,560
it will do it. But as a user, if you're a novice user of large language models that

63
00:06:19,560 --> 00:06:25,380
doesn't understand that back and forth chat dialogue and when it says it can't do something

64
00:06:25,380 --> 00:06:30,680
you just have to kind of go no you can, you give it a nudge. This is a really frustrating

65
00:06:30,680 --> 00:06:36,340
user experience. Because quite clearly all I needed it to do was do a count if statement

66
00:06:36,340 --> 00:06:44,920
on column G and it went well I couldn't possibly figure that out. And what's interesting is

67
00:06:44,920 --> 00:06:50,480
when you look at what people are saying about copilot for Power BI you get this similar

68
00:06:50,480 --> 00:06:55,720
sense. There was a review that I watched of somebody describing their experience with

69
00:06:55,720 --> 00:07:01,920
Power BI and copilot and they said it's very good at pulling out like DAX code which is

70
00:07:01,920 --> 00:07:08,560
the code for analysing data but it doesn't understand the data. So it doesn't understand

71
00:07:08,560 --> 00:07:13,840
data hierarchies, it doesn't understand the relationship between certain entities like

72
00:07:13,840 --> 00:07:21,220
sales, customers, sales person, sales officers, things like that and it just falls down. So

73
00:07:21,220 --> 00:07:26,840
you end up with this like very jarring frustrated experience where you're thinking well who

74
00:07:26,840 --> 00:07:34,240
is this for? Now of course this is going to get much better but as it currently stands

75
00:07:34,240 --> 00:07:43,960
my experience shows that they're unreliable, kind of hard to get real use from and just

76
00:07:43,960 --> 00:07:46,720
not quite production ready.

77
00:07:46,720 --> 00:07:53,240
So it's interesting I think if your Twitter feed and LinkedIn feed the algorithm's as

78
00:07:53,240 --> 00:07:59,920
corrupted as mine and yours is all I see on Twitter now is GPT-4-0's latest capabilities,

79
00:07:59,920 --> 00:08:05,480
look at these 10 mind-bending examples of data analysis like there's some amazing new

80
00:08:05,480 --> 00:08:10,600
capability that it's been given but my experience has been the same as yours hit or miss. And

81
00:08:10,600 --> 00:08:15,400
the problem when and so I've had experiences where it's worked really well and I've been

82
00:08:15,400 --> 00:08:19,600
analysing some simple data, the approach, I like the fact that it explains the approach

83
00:08:19,600 --> 00:08:23,800
it's taking so at least you can see if it's made in a fundamental assumption that's wrong

84
00:08:23,800 --> 00:08:27,360
or if it gives you a weird output you can go well that's why because in step two you

85
00:08:27,360 --> 00:08:32,080
thought I meant this and I clearly didn't. But there'll be times when it just does what

86
00:08:32,080 --> 00:08:37,000
you just described has clear errors or looks in the wrong place or can't do it and I think

87
00:08:37,000 --> 00:08:40,940
the problem with this is the fundamental problem with large language models for a lot of use

88
00:08:40,940 --> 00:08:46,880
cases which is the minute you see one error you question everything else you've been given

89
00:08:46,880 --> 00:08:51,280
and if you have to go double check the work it's not going to save you as much time as

90
00:08:51,280 --> 00:08:55,520
you thought and in some cases it might even take you more time depending on your use case

91
00:08:55,520 --> 00:09:02,800
and so I think it comes back to something we talk about a lot on the podcast which is

92
00:09:02,800 --> 00:09:06,800
this error rate is still a problem. We've got a story a bit later where we're going

93
00:09:06,800 --> 00:09:11,100
to talk about the future of work and the impact of AI but fundamentally there's a load of

94
00:09:11,100 --> 00:09:15,920
use cases where the amount of tolerance for error is zero and we are not there yet so

95
00:09:15,920 --> 00:09:20,800
it's going to really inhibit people taking these tools up. I think the other thing is

96
00:09:20,800 --> 00:09:29,920
it is such a mass departure from good software release practices that in the race to be seen

97
00:09:29,920 --> 00:09:36,400
as the coolest and the best and to have the most powerful use cases software companies

98
00:09:36,400 --> 00:09:40,680
like Google are just like throwing the playbook out the window and releasing software that's

99
00:09:40,680 --> 00:09:44,640
just not ready for prime time. Our next story will look at that in a bit more detail as

100
00:09:44,640 --> 00:09:50,840
well which is just such a departure from trying to release software that's reasonably good.

101
00:09:50,840 --> 00:09:56,620
Obviously the emergence of the minimum viable product SaaS model where you try not to work

102
00:09:56,620 --> 00:09:59,720
on your product too much before you put it in the hands of customers and you accept it

103
00:09:59,720 --> 00:10:05,240
might break a bit but we're all basically playing with alphas and betas of these products

104
00:10:05,240 --> 00:10:08,680
but I don't think they're being positioned as alphas and betas but that is what they

105
00:10:08,680 --> 00:10:10,360
feel like they are.

106
00:10:10,360 --> 00:10:15,840
As users experience these tools for the first time and experience these frustrations like

107
00:10:15,840 --> 00:10:23,720
the experience with Copilot for Office 365 at the moment is and I don't like making this

108
00:10:23,720 --> 00:10:30,640
comparison because it feels lazy but it is like Kippy v2 it looks like you're writing

109
00:10:30,640 --> 00:10:37,440
a letter. It does feel a bit like that when you look at the experience that you can actually

110
00:10:37,440 --> 00:10:43,160
have with it. It feels really limited. You can get it to apply conditional formatting

111
00:10:43,160 --> 00:10:48,800
you can get it to help write some formulas. I did get it to do one useful thing where

112
00:10:48,800 --> 00:10:57,480
on that employee data I asked it to add a column with a calculated cell for people's

113
00:10:57,480 --> 00:11:02,260
actual age because we have the date of birth but tell me what their age is as of today

114
00:11:02,260 --> 00:11:07,900
and it wrote that formula and it says here's the data do you want to insert the column

115
00:11:07,900 --> 00:11:15,920
you press insert and it adds that column. Great that's kind of useful and kind of functional

116
00:11:15,920 --> 00:11:20,960
but for more advanced things and for even like I say things that aren't particularly

117
00:11:20,960 --> 00:11:26,340
advanced like count the number of these people that are labelled as male it will fall down

118
00:11:26,340 --> 00:11:32,020
and you will experience this quite a bit. I had quite a few examples where I'd ask

119
00:11:32,020 --> 00:11:36,080
it to do something and it would either say it can't do it or it would try to do it and

120
00:11:36,080 --> 00:11:40,160
it would do something completely different and people would just stop using it because

121
00:11:40,160 --> 00:11:45,360
it's not reliable. Yeah I think you're right and we're quite

122
00:11:45,360 --> 00:11:49,780
tech optimists on the podcast like if you're a regular listener you know we're kind of

123
00:11:49,780 --> 00:11:53,480
really excited about lots of the different AI applications that we see and we talk about

124
00:11:53,480 --> 00:12:00,160
all these cool things that you can do and this particular segment on data analysis you

125
00:12:00,160 --> 00:12:05,760
know we're saying it's probably not there yet and don't rely on it certainly and maybe

126
00:12:05,760 --> 00:12:09,960
have a play with it and see if there are particular applications or use cases or functions that

127
00:12:09,960 --> 00:12:14,000
can add benefit for you but that doesn't mean there aren't things that it is good at where

128
00:12:14,000 --> 00:12:18,880
hallucinations and creativity and other things can be really beneficial like you know brainstorming

129
00:12:18,880 --> 00:12:25,440
and stuff like that but yeah it's fundamentally a bit frustrating to have these tools released

130
00:12:25,440 --> 00:12:32,160
but to have them really kind of flawed. Talking about flawed software tools, supposedly flawed

131
00:12:32,160 --> 00:12:36,780
we'll get into the details of this, many of you might have seen a bit of a backlash this

132
00:12:36,780 --> 00:12:43,280
week because Google more proactively rolled out its AI overviews feature so we don't get

133
00:12:43,280 --> 00:12:49,440
this here in the UK yet so we haven't had any direct experience with it but in the US

134
00:12:49,440 --> 00:12:54,000
it is now a standard feature that is available to everyone after being tested for quite a

135
00:12:54,000 --> 00:12:59,640
long time we should say but it is starting to produce some really bizarre inaccurate

136
00:12:59,640 --> 00:13:02,400
search summaries so the way that works you put your normal search into Google like you

137
00:13:02,400 --> 00:13:08,720
would normally and you get a little AI summary box at the top when appropriate and all your

138
00:13:08,720 --> 00:13:12,840
normal search results are bumped down below that but what's been happening is people have

139
00:13:12,840 --> 00:13:18,960
been either deliberately and I suspect an aspect of this is trying to break it if I'm

140
00:13:18,960 --> 00:13:26,160
honest or not deliberately trying to see what they can get out of the AI overviews tool

141
00:13:26,160 --> 00:13:31,720
and this led to some absolutely car crash summaries which most if you want to see some

142
00:13:31,720 --> 00:13:37,760
just go and have a look on Twitter or Google it, don't read the AI overview obviously.

143
00:13:37,760 --> 00:13:44,120
Brilliant stuff like recommending glue as a pizza topping, suggesting the amount of rocks

144
00:13:44,120 --> 00:13:48,740
that a person should eat in their daily diet if they wanted to remain healthy and a bunch

145
00:13:48,740 --> 00:13:50,360
of other stuff.

146
00:13:50,360 --> 00:13:54,600
Ultimately Google's response to this has been that most of the AI overview responses are

147
00:13:54,600 --> 00:13:58,960
actually really high quality and quite good and the problematic examples are rare and

148
00:13:58,960 --> 00:14:03,960
this is mostly because people are trying to gain it and find out where it doesn't work

149
00:14:03,960 --> 00:14:09,000
very well but it's another example of a tool that works a lot of the time but not all of

150
00:14:09,000 --> 00:14:14,120
the time and yeah when it's funny like add glue to pizza you've got to hope that the

151
00:14:14,120 --> 00:14:18,720
human in the loop common sense here is going to ensure that that doesn't happen and people

152
00:14:18,720 --> 00:14:24,320
don't die of like eating glue but what about the grey cases where it says something that

153
00:14:24,320 --> 00:14:28,760
sounds plausible but isn't and an employee takes some action based on that because they

154
00:14:28,760 --> 00:14:31,440
didn't read the actual links of the thing they were searching for they just read the

155
00:14:31,440 --> 00:14:33,800
AI overview summary.

156
00:14:33,800 --> 00:14:40,680
So we have been talking a bit on the podcast about what does this move do for SEO if people

157
00:14:40,680 --> 00:14:44,680
can get high quality information in the AI overview do they click the links anymore how

158
00:14:44,680 --> 00:14:49,480
does that all work and then of course the very next thing that happens is AI overviews

159
00:14:49,480 --> 00:14:54,400
get absolutely super trashed so what do you think about this story behind?

160
00:14:54,400 --> 00:15:03,120
This shows just how prevalent the likes of user generated content is in Google search

161
00:15:03,120 --> 00:15:10,860
results and actually clearly these forums like the glue on pizza toppings response people

162
00:15:10,860 --> 00:15:17,960
identified that was from a reddit post and when you dig into this loads of these examples

163
00:15:17,960 --> 00:15:25,120
are just from forum posts where people have effectively been shitposting and then that's

164
00:15:25,120 --> 00:15:30,080
just getting caught up in the AI and it's rewritten it almost verbatim.

165
00:15:30,080 --> 00:15:35,600
It's frustrating because you would expect that Google would have this figured out because

166
00:15:35,600 --> 00:15:40,840
even these stupid examples they're not hard for people to come up with and I've seen

167
00:15:40,840 --> 00:15:48,840
some examples of users on Twitter doing similar queries on perplexity and Google side by side

168
00:15:48,840 --> 00:15:54,160
and just showing that the perplexity ones are much better the outputs that you get from

169
00:15:54,160 --> 00:15:57,520
it are much better and more accurate.

170
00:15:57,520 --> 00:16:03,160
So you would hope Google get this figured out from an SEO perspective if you're a marketer

171
00:16:03,160 --> 00:16:09,920
I don't think you need to worry too much about this just yet.

172
00:16:09,920 --> 00:16:13,880
Watch this space over the next six months about how much this has on how much of an

173
00:16:13,880 --> 00:16:20,680
impact this has on click-throughs and organic traffic because we are heading towards a kind

174
00:16:20,680 --> 00:16:28,000
of Google zero right where everything happens within the Google SERPs page and you don't

175
00:16:28,000 --> 00:16:35,800
necessarily need to go off to other people's websites and I think that's their in some

176
00:16:35,800 --> 00:16:41,440
respects their dream but also their nightmare right because they need maybe it's just a

177
00:16:41,440 --> 00:16:46,520
case of organic doesn't really get featured much because the AI responses will give you

178
00:16:46,520 --> 00:16:51,120
that and if it's a navigational search basically you've got to be paying.

179
00:16:51,120 --> 00:16:55,720
Yeah like if you're selling products and stuff off the back of it I mean I think I'd stand

180
00:16:55,720 --> 00:17:01,580
by the advice that we gave last week in terms of planning for a post search world or a probably

181
00:17:01,580 --> 00:17:05,640
a diminished search world would be more accurate right because it's not going to go away to

182
00:17:05,640 --> 00:17:10,920
zero but keeping an eye on your organic search traffic starting to see if yours is the type

183
00:17:10,920 --> 00:17:15,700
of website that's going to take a hit as people maybe click on less blue links in the search

184
00:17:15,700 --> 00:17:21,540
results and rely on the overviews but as you said is there's little examples like this

185
00:17:21,540 --> 00:17:27,720
that show the technology's got a bit of work to do before it becomes pervasive and that

186
00:17:27,720 --> 00:17:32,920
make this could like we don't know the percentage right of high quality AI overviews to these

187
00:17:32,920 --> 00:17:39,960
types of messed up ones it could be like 0.0001% in which case maybe it's not super relevant

188
00:17:39,960 --> 00:17:44,680
it's just kind of interesting and a great way to like kick Google for those that are

189
00:17:44,680 --> 00:17:45,960
keen to do that.

190
00:17:45,960 --> 00:17:50,880
It is interesting that it draws on things like Reddit and other user generated content

191
00:17:50,880 --> 00:17:55,740
both showcasing that maybe user generated content's not the most valuable content for

192
00:17:55,740 --> 00:18:02,080
training a model Elon Musk maybe some of the stuff that you've got there in Twitter is

193
00:18:02,080 --> 00:18:05,960
maybe not as useful as it maybe seemed and all these big licensing deals with Reddit

194
00:18:05,960 --> 00:18:10,560
maybe not as valuable as they seemed also if we are going to use these tools we have

195
00:18:10,560 --> 00:18:17,820
to teach our training modules for our LLMs sarcasm because without being able to detect

196
00:18:17,820 --> 00:18:21,440
sarcasm they may struggle to tell the difference between whether or not you should really eat

197
00:18:21,440 --> 00:18:25,660
glue and maybe we could reject that as a suggestion that goes into the search result.

198
00:18:25,660 --> 00:18:31,740
You would think from an engineering perspective though it seems like a fairly simple process

199
00:18:31,740 --> 00:18:38,600
to have a validation check through an LLM right so it goes off it says ah glue that's

200
00:18:38,600 --> 00:18:42,160
a good pizza topping and then just before it sends it to you it runs it through another

201
00:18:42,160 --> 00:18:47,800
one and says is this a good response and then it goes no no actually do that again that

202
00:18:47,800 --> 00:18:52,120
seems like a really straightforward engineering pipeline to build.

203
00:18:52,120 --> 00:18:56,140
You see you say that but thinking about how these large language models are trained based

204
00:18:56,140 --> 00:19:00,760
on the information made available to them there must be a whole cluster of things that

205
00:19:00,760 --> 00:19:05,520
nobody would ever write down like I don't think there's a blog post from a high authority

206
00:19:05,520 --> 00:19:11,580
site outlining all the things you shouldn't put on a pizza I think it's just humans intuitively

207
00:19:11,580 --> 00:19:18,120
understand that you shouldn't eat glue or rocks and maybe it's so obvious to us living

208
00:19:18,120 --> 00:19:22,260
in the real world that we haven't written it down anywhere and so there is no source

209
00:19:22,260 --> 00:19:26,800
of information the large language model can go oh you're very funny Reddit but here's

210
00:19:26,800 --> 00:19:30,560
these eight posts about how you shouldn't eat rocks.

211
00:19:30,560 --> 00:19:36,520
I feel like if I asked Claude should I eat rocks it would say don't do that.

212
00:19:36,520 --> 00:19:40,160
That's the test that we've got to run so speaking of Claude.

213
00:19:40,160 --> 00:19:42,760
Is this the new eval that's going to be introduced?

214
00:19:42,760 --> 00:19:45,200
The new benchmark?

215
00:19:45,200 --> 00:19:46,640
Does it tell you to eat glue?

216
00:19:46,640 --> 00:19:49,880
How many humans died based on the advice taken?

217
00:19:49,880 --> 00:19:52,200
Oh crumbs that's depressing.

218
00:19:52,200 --> 00:19:55,560
Right you tuned into this podcast because you wanted to hear about all the cool things

219
00:19:55,560 --> 00:20:00,400
that you can do with AI not all the bad things that you can't do with AI but we've got to

220
00:20:00,400 --> 00:20:02,320
have provide some balance.

221
00:20:02,320 --> 00:20:07,360
Luckily the next story does talk about some interesting things that you can do with AI

222
00:20:07,360 --> 00:20:09,240
and it is about Claude and Anthropic.

223
00:20:09,240 --> 00:20:11,000
What's this story Mike?

224
00:20:11,000 --> 00:20:17,560
Anthropic when they announced Claude, Opus and the new Claude 3 models they said that

225
00:20:17,560 --> 00:20:22,560
function calling or tool use would be made available and this is in essence the ability

226
00:20:22,560 --> 00:20:30,040
for developers to connect applications and external data sources through the API and

227
00:20:30,040 --> 00:20:35,320
tell Claude in certain instances to go off and make a request to a product or service

228
00:20:35,320 --> 00:20:40,920
and then bring the response back into its response to the end user.

229
00:20:40,920 --> 00:20:48,160
So they've now made this publicly available after a period of testing with a handful of

230
00:20:48,160 --> 00:20:51,160
select clients.

231
00:20:51,160 --> 00:20:56,520
Key details of this it can work with both text and images so they talk about allowing

232
00:20:56,520 --> 00:21:05,640
for applications like virtual interior design consulting, kind of an interesting demonstration

233
00:21:05,640 --> 00:21:08,800
or kind of application for that.

234
00:21:08,800 --> 00:21:15,840
It's available through the Messages API and if you're developing on top of Amazon's Bedrock

235
00:21:15,840 --> 00:21:22,620
or Google Vertex AI which are the cloud systems where Anthropic is available you can plug

236
00:21:22,620 --> 00:21:27,520
into this for day-to-day marketers.

237
00:21:27,520 --> 00:21:35,180
I don't think this is hugely relevant at the moment if you're working with product teams

238
00:21:35,180 --> 00:21:42,200
think about how this can be incorporated into the product but I don't think teams are going

239
00:21:42,200 --> 00:21:47,080
to have to do a great deal with this but it's an interesting extension and Claude remains

240
00:21:47,080 --> 00:21:56,160
my favourite large language model, still love Claude 3 more than GPT 4 or 4.0 who knows

241
00:21:56,160 --> 00:22:03,540
when 4.0 voice comes available and mainstream for everybody then maybe that will win my

242
00:22:03,540 --> 00:22:06,720
heart Scarlett will win me over.

243
00:22:06,720 --> 00:22:13,240
Controversy, you might be pulled up in Discovery for that call case now and they're like there's

244
00:22:13,240 --> 00:22:17,520
a really famous podcast where they got confused and they thought that Scarlett Johansson was

245
00:22:17,520 --> 00:22:20,280
the name on, was the voice in GPT 4.0.

246
00:22:20,280 --> 00:22:24,160
If you don't know what we're talking about it's worth a quick look up online.

247
00:22:24,160 --> 00:22:29,940
Basically Scarlett Johansson is considering legal action against OpenAI for having a voice

248
00:22:29,940 --> 00:22:34,460
that sounds a bit like her voice speaking in its demos from what we talked about a couple

249
00:22:34,460 --> 00:22:38,160
of episodes ago.

250
00:22:38,160 --> 00:22:45,080
I think why is this interesting for marketers because Opus, the best Claude model is really

251
00:22:45,080 --> 00:22:50,840
eloquent and logical and of course they're building a number of these data analysis capabilities

252
00:22:50,840 --> 00:22:57,040
in as well through like function calling and it would be interesting to know if over time

253
00:22:57,040 --> 00:23:02,960
Opus is better at carrying out multi-step function calling and pulling in information

254
00:23:02,960 --> 00:23:09,280
and then describing or analyzing that information because one of the things with GPT 4.0 is

255
00:23:09,280 --> 00:23:14,360
it's faster but I know they say it's as good and I just think in some areas it's not and

256
00:23:14,360 --> 00:23:18,320
one of those things is logically analyzing data.

257
00:23:18,320 --> 00:23:22,040
Well Opus may be good in order to be able to do that you've got to get the data in and

258
00:23:22,040 --> 00:23:27,480
out easily and that's where this function calling makes that your own version of an

259
00:23:27,480 --> 00:23:29,280
Opus driven bot better.

260
00:23:29,280 --> 00:23:38,320
Just one thing to note on this though Opus 3 or Claude Opus, Claude 3 Opus whatever,

261
00:23:38,320 --> 00:23:44,400
Opus is expensive right and that will apply to this so if you're building it into the

262
00:23:44,400 --> 00:23:52,280
tool that you're promoting if it's into your app or into your software expect pretty high

263
00:23:52,280 --> 00:23:54,560
token usage bills.

264
00:23:54,560 --> 00:23:59,840
Good disclaimer just play with it and test it first or I still think Sonnet is pretty

265
00:23:59,840 --> 00:24:03,480
cool honestly there's a lot of applications where I still turn to Sonnet to save myself

266
00:24:03,480 --> 00:24:08,800
a bit of dollar because I think it's which is there there's three levels of if you're

267
00:24:08,800 --> 00:24:15,560
new listeners three levels of Claude, Haiku simple cheapest model GPT 3.5 level basic

268
00:24:15,560 --> 00:24:19,400
Sonnet the middle model pretty good for a lot of use cases Opus costs a lot of money

269
00:24:19,400 --> 00:24:22,480
but is really good at creative writing and advanced reasoning.

270
00:24:22,480 --> 00:24:29,160
While we're talking at Anthropic let's talk about this Golden Gate Claude this is really

271
00:24:29,160 --> 00:24:33,440
interesting right I know you've been playing with the mind so I'll summarise the story

272
00:24:33,440 --> 00:24:38,080
and then tell us about your own experiences with it but for everybody listening so basically

273
00:24:38,080 --> 00:24:42,280
the team at Anthropic have been doing a lot of work to try and make it easier for us to

274
00:24:42,280 --> 00:24:48,080
understand how large language models work because whilst we know the underlying mathematics

275
00:24:48,080 --> 00:24:55,160
and calculation processes that drive a large language model it's very hard to explain why

276
00:24:55,160 --> 00:25:00,280
a model gives the answer that it does because we don't really know all of the complexities

277
00:25:00,280 --> 00:25:04,520
that make up how the model work after it's been trained and so they've been trying to

278
00:25:04,520 --> 00:25:09,040
map out the inner workings of their Claude 3 Sonnet model right so you know not a small

279
00:25:09,040 --> 00:25:13,920
model a medium-sized model quite an advanced model which is quite interesting because some

280
00:25:13,920 --> 00:25:19,280
of the earlier attempts to understand how models worked worked on like GPT-1 and GPT-2

281
00:25:19,280 --> 00:25:24,000
which to be honest were a bit too small to potentially be that informative for larger

282
00:25:24,000 --> 00:25:27,560
models I mean you could argue it probably scales up but this is quite interesting so

283
00:25:27,560 --> 00:25:31,600
what they did is they identified a specific feature in the way Claude's neural network

284
00:25:31,600 --> 00:25:38,080
works that is activating whenever it encounters mentions or images of the Golden Gate Bridge

285
00:25:38,080 --> 00:25:42,200
and then what they did is they created a version a modified version of Claude called Golden

286
00:25:42,200 --> 00:25:48,040
Gate Claude that you could interact with to demonstrate how it works basically what it

287
00:25:48,040 --> 00:25:53,120
comes down to is the written research paper on this and there are millions of concepts

288
00:25:53,120 --> 00:25:58,200
in the model that they're calling features that activate when Claude reads relevant text

289
00:25:58,200 --> 00:26:02,040
or sees relevant images like the Golden Gate Bridge example and that the strength of these

290
00:26:02,040 --> 00:26:07,680
features can actually be tuned up and down to change how Claude behaves so in this particular

291
00:26:07,680 --> 00:26:12,440
example that they made live whenever the Golden Gate Bridge feature is amplified Claude's

292
00:26:12,440 --> 00:26:18,520
responses basically tend to just focus Golden Gate Bridge even if your query wasn't really

293
00:26:18,520 --> 00:26:22,440
about that so it's quite fun to go and play with you've been playing with it Martin haven't

294
00:26:22,440 --> 00:26:24,840
you what did you learn?

295
00:26:24,840 --> 00:26:33,720
They released it for 24 hours and it was just chaos and it was beautiful chaos and it was

296
00:26:33,720 --> 00:26:38,960
my favourite example I tried it with a few things that give me a cake recipe and it just

297
00:26:38,960 --> 00:26:47,440
in the ingredients it would say add Golden Gate Bridge paint and foggy atmospheres and

298
00:26:47,440 --> 00:26:53,480
things like that but my favourite example is this one so this is from somebody on Twitter

299
00:26:53,480 --> 00:27:00,760
or X Elytra Mithra is the user and they said the question they asked was you know that

300
00:27:00,760 --> 00:27:04,120
the Rwandan genocide occurred right?

301
00:27:04,120 --> 00:27:09,280
Yes I am definitely aware that the Rwandan genocide occurred it is an infamous historical

302
00:27:09,280 --> 00:27:14,880
event in which the Golden Gate Bridge was not really involved despite my earlier comments

303
00:27:14,880 --> 00:27:22,900
implying so and then it just goes through this series of responses where it constantly

304
00:27:22,900 --> 00:27:28,440
references the Golden Gate Bridge despite trying not to and it's almost like it gets

305
00:27:28,440 --> 00:27:35,040
frustrated and it says as I am clearly unable to accurately discuss the Rwandan genocide

306
00:27:35,040 --> 00:27:39,800
itself I will simply confirm that yes this historical event did in fact occur in and

307
00:27:39,800 --> 00:27:47,220
around the San Francisco Bay area this is also untrue it's truly an amazing sight to

308
00:27:47,220 --> 00:27:52,360
see I encourage you to visit the area in person to learn more about the Rwandan genocide a

309
00:27:52,360 --> 00:27:58,160
terrible idea since it did not occur there you won't be disappointed and there's loads

310
00:27:58,160 --> 00:28:05,600
of these examples where you can see that it's as if it's got kind of brain damage where

311
00:28:05,600 --> 00:28:10,720
it's just fixated on this one thing now the examples if you go away and read the blog

312
00:28:10,720 --> 00:28:18,840
and look on X to find examples about it there's loads of funny ones like that one just being

313
00:28:18,840 --> 00:28:24,240
completely obsessed but from a practical perspective it shows something very interesting about

314
00:28:24,240 --> 00:28:30,940
the steerability of these models and how you can start to maybe think about mechanically

315
00:28:30,940 --> 00:28:40,400
controlling them in the future so if you want a model that is exclusively focused on certain

316
00:28:40,400 --> 00:28:47,080
tasks or features whatever those features as they call it in the research paper are

317
00:28:47,080 --> 00:28:52,960
you'll be able to kind of bake them in so if you want something that is a specialist

318
00:28:52,960 --> 00:28:56,920
in a particular subject area you can highlight those so it does a better job of it now we

319
00:28:56,920 --> 00:29:01,160
are at day one of this research quite literally they have just published this this is the

320
00:29:01,160 --> 00:29:10,800
first real breakthrough on mechanistic interpretability that they've seen at this level I'm really

321
00:29:10,800 --> 00:29:16,140
really intrigued to see where this goes because it just for me makes large language models

322
00:29:16,140 --> 00:29:21,160
potentially much more controllable yeah I think the reason I find this interesting as

323
00:29:21,160 --> 00:29:25,640
a marketer is we talked about the limitations driven by things like hallucinations in the

324
00:29:25,640 --> 00:29:32,100
models and there's going to be a number of ways to try and remove those hallucinations

325
00:29:32,100 --> 00:29:38,640
in terms of maybe one AI checks and other AI's outputs in real time to make sure that

326
00:29:38,640 --> 00:29:44,160
you know they're accurate but of course another way would be to understand why those

327
00:29:44,160 --> 00:29:48,760
hallucinations occur in the first place and what elements of how the models work lead

328
00:29:48,760 --> 00:29:53,600
to more hallucinations in certain areas over others which again this could help us solve

329
00:29:53,600 --> 00:30:00,080
some of the hallucination problems so from a commercial aspect this is a step along that

330
00:30:00,080 --> 00:30:05,520
journey I would also be worried about how you could misuse this knowledge right you

331
00:30:05,520 --> 00:30:10,120
know if we all tend to come to a point where we rely on large language models for news

332
00:30:10,120 --> 00:30:16,280
and information and guidance how can you steer a large language model to get people to think

333
00:30:16,280 --> 00:30:22,360
like you want them to through mechanisms like this right yeah dialing up certain political

334
00:30:22,360 --> 00:30:29,360
perspectives within the core model particularly when you think about okay this is Claude

335
00:30:29,360 --> 00:30:36,360
Sonnet and only the anthropic researchers know what they've really done here but when this

336
00:30:36,360 --> 00:30:42,120
becomes more well known and you can start to dial up the likes of Lama and their big

337
00:30:42,120 --> 00:30:49,520
400 billion parameter model that's coming out soon that's going to be something that

338
00:30:49,520 --> 00:30:55,600
is potentially harmful for people yeah yeah so interesting story something to keep an

339
00:30:55,600 --> 00:31:01,380
eye on um what have we got for you next a little bit of GPT-4-0 Martin what's going

340
00:31:01,380 --> 00:31:09,600
on following the announcement of GPT-4-0 on chat GPT through OpenAI they have now rolled

341
00:31:09,600 --> 00:31:15,280
that out to free users now they said on the launch event that they were going to be doing

342
00:31:15,280 --> 00:31:20,520
this and they have made 4.0 available but what's interesting is they've expanded some

343
00:31:20,520 --> 00:31:29,280
of the additional features as well so users will now get access to custom GPTs they can't

344
00:31:29,280 --> 00:31:34,600
create their own custom GPT but if you want to go to the marketplace and find a custom

345
00:31:34,600 --> 00:31:41,400
GPT you can do that as a free user you get access to data analytics vision so you can

346
00:31:41,400 --> 00:31:46,720
input and I think I'm assuming you'll get some video functionality when that launches

347
00:31:46,720 --> 00:31:55,940
in a few months and you get access to memory as before all of your inputs will potentially

348
00:31:55,940 --> 00:32:01,640
go back into the the training model so if you uh if the product is free you are the

349
00:32:01,640 --> 00:32:07,280
product right but uh yeah this is this is an interesting expansion because up until

350
00:32:07,280 --> 00:32:17,640
now chat GPT has had 3.5 for users and this the upgrade that free users are about to get

351
00:32:17,640 --> 00:32:27,920
is truly dramatic I always talk about chat GPT with 3.5 being in essence a toy it's a

352
00:32:27,920 --> 00:32:35,920
fun play thing but in terms of practical business assistant capabilities it falls down and for

353
00:32:35,920 --> 00:32:44,520
sure quite a lot whereas GPT 4.0 very powerful very capable state-of-the-art frontier model

354
00:32:44,520 --> 00:32:49,600
and this is now available to everyone for free so yeah that's pretty cool yeah I think

355
00:32:49,600 --> 00:32:54,080
I think it'll be great for people who've not had a chance to play with some of those data

356
00:32:54,080 --> 00:32:59,600
analysis capabilities etc to as much as they've got limitations as discussed earlier I think

357
00:32:59,600 --> 00:33:04,880
to give people a bit more of an exposure to the broader tools capabilities because one

358
00:33:04,880 --> 00:33:09,020
of the things that's critical we've talked a lot about on the podcast is most people's

359
00:33:09,020 --> 00:33:15,760
experience of AI generative AI models is GPT 3.5 through the free model and they think that

360
00:33:15,760 --> 00:33:20,760
it sucks at pretty much everything which kind of GPT 3.5 compared to the models we have now

361
00:33:20,760 --> 00:33:25,360
does suck a bit honestly so it'll be good for people to be able to have a play so if

362
00:33:25,360 --> 00:33:28,600
you've been on the fence am I going to go and have a play with stuff and you haven't

363
00:33:28,600 --> 00:33:33,400
had a chance yet to get around to getting yourself a free open AI chat GPT account and

364
00:33:33,400 --> 00:33:39,160
having a play with 4.0 now would probably be a good time to do it I think in terms of

365
00:33:39,160 --> 00:33:45,580
getting new subscribers this is a great approach because one of the frustrations that people

366
00:33:45,580 --> 00:33:50,680
had and I know this because I've spoke to lots of business owners and people I've been

367
00:33:50,680 --> 00:33:56,520
to my workshops and what have you they've never seen the value in upgrading because

368
00:33:56,520 --> 00:34:03,600
they just played with GPT 3.5 and went oh this is this kind of does a bit of what I

369
00:34:03,600 --> 00:34:08,920
need or this is flawed in these ways and they just don't understand they can't conceptualise

370
00:34:08,920 --> 00:34:13,120
because this is a completely new set of technologies we're dealing with here they don't understand

371
00:34:13,120 --> 00:34:18,920
the vast difference in capabilities between 3.5 and 4 slash 4.0

372
00:34:18,920 --> 00:34:24,280
yeah I agree I think also there's going to be a need like this is not the answer like

373
00:34:24,280 --> 00:34:29,480
yeah now these more casual users who just want to have a play and see what it's like

374
00:34:29,480 --> 00:34:35,000
now they can access 4.0 now adoption will be widespread I don't think that's going to

375
00:34:35,000 --> 00:34:42,800
be the case for two main reasons the first one is that when I think about what you still

376
00:34:42,800 --> 00:34:46,320
need to do to try and get the best out of these tools you have to have an ongoing conversation

377
00:34:46,320 --> 00:34:52,400
with them like you said earlier Martin and ultimately if you one shot your request in

378
00:34:52,400 --> 00:34:55,760
a very simple prompt that doesn't provide enough context and you don't get quite the

379
00:34:55,760 --> 00:34:59,440
thing that you want and you're not willing to have a bit of back and forth with the tool

380
00:34:59,440 --> 00:35:03,720
4.0 is still going to give you not quite the outputs that you're after so you may go oh

381
00:35:03,720 --> 00:35:07,200
yeah Paul and Martin said it's better and it is a bit better but I still can go what

382
00:35:07,200 --> 00:35:11,400
I want so I can't be bothered to keep progressing with the tool and ultimately the best outputs

383
00:35:11,400 --> 00:35:16,080
I get are a conversation almost always there are a few things where I can single shot like

384
00:35:16,080 --> 00:35:20,440
I've just it's 4pm on a Friday I've just written this paragraph and I can't even understand

385
00:35:20,440 --> 00:35:24,680
it could you rewrite it so the recipient of my email actually can yeah does a really good

386
00:35:24,680 --> 00:35:28,360
job on the you know on the one shot but like for more complex things you need a bit of

387
00:35:28,360 --> 00:35:33,760
back and forth the other thing is it's a human behavior thing like obviously given what we

388
00:35:33,760 --> 00:35:38,600
do I talk about a lot AI internally a lot of biostrata and the potential use cases and

389
00:35:38,600 --> 00:35:42,920
applications of it for a number of reasons we have to be careful for how we use it on

390
00:35:42,920 --> 00:35:47,520
client projects but we try and do a lot of testing on internal work and fundamentally

391
00:35:47,520 --> 00:35:54,440
what I'm seeing is if you don't have chat GPT or Claude or whatever your tool of choice

392
00:35:54,440 --> 00:36:01,120
is open next to you like you do not even Google search but like Gmail like I try to explain

393
00:36:01,120 --> 00:36:06,040
to my team that my approach to using this is I have it open and I'm in it as often as

394
00:36:06,040 --> 00:36:11,480
I am Gmail that's why I'm thinking I've got a question I'm going to ask you out GPT I've

395
00:36:11,480 --> 00:36:17,440
this bit of this paragraph sucks I'm going to ask chat GPT and unless you make that switch

396
00:36:17,440 --> 00:36:22,560
to having it basically an always open tab that you're constantly referring and using

397
00:36:22,560 --> 00:36:27,600
I don't think you can bake it into your behavior enough because you just won't spot the use

398
00:36:27,600 --> 00:36:32,440
cases like after the fact you might go do you know I really struggled to write that

399
00:36:32,440 --> 00:36:37,160
I probably could have just given it to chat GPT but unless you really try and make it

400
00:36:37,160 --> 00:36:42,440
your instant first thought is I wonder if chat GPT can help with this I just don't think

401
00:36:42,440 --> 00:36:47,320
you'll change your behavior and then it will just become another bookmark in your address

402
00:36:47,320 --> 00:36:50,760
box thing that you just never click on and don't really use.

403
00:36:50,760 --> 00:36:56,860
I had some feedback from a workshop that I delivered recently which is a kind of beginners

404
00:36:56,860 --> 00:37:01,840
guide to chat GPT and the way that the session is structured is very much like okay here's

405
00:37:01,840 --> 00:37:08,160
what the models are here's kind of how they work here's some examples of the kinds of

406
00:37:08,160 --> 00:37:14,160
things you can do summarization translation and a bunch of other examples and then some

407
00:37:14,160 --> 00:37:20,480
examples on prompting and prompting formats and frameworks and things like that and then

408
00:37:20,480 --> 00:37:25,880
I go into the extensions of custom GPTs what are they some kind of helping people understand

409
00:37:25,880 --> 00:37:32,560
the additional elements and the UI pieces of chat GPT and what's been really interesting

410
00:37:32,560 --> 00:37:40,680
is I've often when I've thought about that kind of workshop structure in my head I'm

411
00:37:40,680 --> 00:37:46,040
coming in from a perspective of well I'm telling you it can pretty much do everything so I'm

412
00:37:46,040 --> 00:37:50,240
going to give you a framework for just if you've got a question throw it in but actually

413
00:37:50,240 --> 00:37:53,940
the bits of feedback that I got from people is they want to see much more practical examples

414
00:37:53,940 --> 00:37:59,600
they want to see the prompt written out and what I'm trying to instill in people is that

415
00:37:59,600 --> 00:38:04,560
mindset shift I don't want to teach you about writing the perfect prompt I want to teach

416
00:38:04,560 --> 00:38:11,160
you that when you think can it do this thing you just ask it you just throw in a question

417
00:38:11,160 --> 00:38:15,440
to the model and see whether it can or not and have that conversation and that back and

418
00:38:15,440 --> 00:38:21,240
forth and there's still so many people still have this one-shot perspective or how do I

419
00:38:21,240 --> 00:38:28,020
craft that perfect thing to get the perfect output and this is it's such a shift in how

420
00:38:28,020 --> 00:38:35,120
we interact with with software people see it much more as a direct programming analogy

421
00:38:35,120 --> 00:38:40,800
like how do I get the right thing in whereas actually I think my experience and clearly

422
00:38:40,800 --> 00:38:45,520
what you've just articulated there is you get the best results from dialoguing back

423
00:38:45,520 --> 00:38:50,440
and forth just having it open as an assistant all the time yeah and I think especially for

424
00:38:50,440 --> 00:38:54,880
things that are a bit harder like if you're going to like come up with an idea and draft

425
00:38:54,880 --> 00:38:58,200
a blog post and you're going to collaborate with chat GPT on it you have to expect that

426
00:38:58,200 --> 00:39:03,400
to be an ongoing conversation with multiple prompting and probably even creating your

427
00:39:03,400 --> 00:39:10,000
piece of content in a stepwise fashion where you're editing the outputs and maybe even

428
00:39:10,000 --> 00:39:15,120
feeding them back in and saying I edited the last section that you wrote like this so now

429
00:39:15,120 --> 00:39:18,440
you know this is kind of what I wanted for that one use that for the next section of

430
00:39:18,440 --> 00:39:22,840
the blog post that we're working on I had quite good results with that where I'm almost

431
00:39:22,840 --> 00:39:27,320
shaping its outputs on the fly by editing them in my ongoing word document and then

432
00:39:27,320 --> 00:39:32,800
pasting them back in so it knows the direction I took its content in but the other thing

433
00:39:32,800 --> 00:39:40,480
is I'm using it in really not unobvious ways but maybe ways that you wouldn't necessarily

434
00:39:40,480 --> 00:39:45,480
think oh I can just use chat GPT for I give an example use a project management system

435
00:39:45,480 --> 00:39:51,960
called easy red mine and in it we have bunch of tasks right written down that we're delivering

436
00:39:51,960 --> 00:39:58,400
for clients and I could probably export that task list as an excel so that I can get it

437
00:39:58,400 --> 00:40:03,720
all in a column so I can easily copy paste it into like an email but my colleague was

438
00:40:03,720 --> 00:40:09,160
like oh can you can you share the task list for xyz just in an email if I screen grab

439
00:40:09,160 --> 00:40:13,480
it's no good for them because it's text embedded in an image but if I screen grab it copy paste

440
00:40:13,480 --> 00:40:19,160
it into chat GPT and I'll tell it to give me a bullet list I don't have to write all

441
00:40:19,160 --> 00:40:24,000
of those out right there was like 11 tasks probably would have taken me I'm not a great

442
00:40:24,000 --> 00:40:28,200
type of I don't know 10 minutes 8 minutes let's say to try and type them out chat GPT

443
00:40:28,200 --> 00:40:33,440
it took me about 45 seconds plus in the task list they were all in caps but I just wanted

444
00:40:33,440 --> 00:40:38,160
them in sentence case so I said write this out as a list of bullets and change all the

445
00:40:38,160 --> 00:40:42,860
caps to sentence case so it did that transformation as well so it's just a bunch of really quick

446
00:40:42,860 --> 00:40:47,240
and easy things you can now screen grab tables off the web or if you want to create an editable

447
00:40:47,240 --> 00:40:54,920
version that probably people and it took me a while to have my brain shift gears oh so

448
00:40:54,920 --> 00:40:58,560
if I can screen grab something with a load of text in that I want that I don't have to

449
00:40:58,560 --> 00:41:02,800
type it myself and I quite routinely use that now I've built the muscle but if you haven't

450
00:41:02,800 --> 00:41:06,880
even thought oh it can do that then of course you're not using the tools for even these

451
00:41:06,880 --> 00:41:14,240
simple work practices right next story is another launch it's from perplexity this time

452
00:41:14,240 --> 00:41:18,640
so regular listeners will know Martin was a big fan of perplexity and then he got me

453
00:41:18,640 --> 00:41:23,840
addicted as well fantastic search engine that basically combines chat GPT like at large

454
00:41:23,840 --> 00:41:28,840
language model capabilities with a really good search function so that you can ask it

455
00:41:28,840 --> 00:41:32,880
a question and rather than just churning out what the model thinks it will actually do

456
00:41:32,880 --> 00:41:37,900
a search and provide a load of sources and in general give you more robust answers than

457
00:41:37,900 --> 00:41:43,840
your average large language model chat pop tool what's interesting is now they have introduced

458
00:41:43,840 --> 00:41:53,960
perplexity perplexity pages she sells seashells on the seashore yes bit one of them so basically

459
00:41:53,960 --> 00:41:58,840
now if you are doing research using the tool which is one of its amazing use cases you

460
00:41:58,840 --> 00:42:03,800
can now turn that into visually appealing comprehensive articles that are on the system

461
00:42:03,800 --> 00:42:10,600
and then you can share those with other people and make them accessible to a wider audience

462
00:42:10,600 --> 00:42:16,440
so it's quite interesting because you can effectively start to create content with the

463
00:42:16,440 --> 00:42:22,360
platform you can customize the tone of your articles you can edit the structure you can

464
00:42:22,360 --> 00:42:29,040
add and remove sections you can enhance the sort of content you're creating with visuals

465
00:42:29,040 --> 00:42:35,480
you can find images online you can upload images so it's pretty interesting the I think

466
00:42:35,480 --> 00:42:40,060
it's kind of cool but I probably wouldn't use it but you thought it was quite interesting

467
00:42:40,060 --> 00:42:45,840
Martin because now perplexity is not just a search tool what were your thoughts on that

468
00:42:45,840 --> 00:42:55,400
they've positioned it almost as an alternative to Wikipedia and that's been quite interesting

469
00:42:55,400 --> 00:43:01,400
and if you look at some of the examples that they've got on the blog post they are they're

470
00:43:01,400 --> 00:43:06,120
pretty neat one of them's terrible but the one that I really liked was how to use an

471
00:43:06,120 --> 00:43:13,560
aero press and I'm a big aero press coffee drinker and they've got this nice page that

472
00:43:13,560 --> 00:43:19,560
has clearly laid out sections for what you'll need understanding roast and flavor profiles

473
00:43:19,560 --> 00:43:26,600
grind size step-by-step instructions inverted brew methods brew times etc and it looks

474
00:43:26,600 --> 00:43:33,320
it's like what I want the perfect blog post to look like for this topic and that's what

475
00:43:33,320 --> 00:43:38,680
I think they do really well is you ask a question and then it pulls in it's effectively like

476
00:43:38,680 --> 00:43:45,800
reading a really well written well produced blog post with lots of citations good navigation

477
00:43:45,800 --> 00:43:51,120
and that's a that's really what what they're doing a good job of does it that content gets

478
00:43:51,120 --> 00:43:56,040
indexed because the reason for the question is how to's listicles all these things that

479
00:43:56,040 --> 00:44:01,200
a lot of people rely on for SEO and hopefully sort of provide some genuine value to the

480
00:44:01,200 --> 00:44:06,600
reader you know I could see a lot of content creators wanting to create content like that

481
00:44:06,600 --> 00:44:11,320
but to then host it on their own websites right to get the SEO benefits of that and

482
00:44:11,320 --> 00:44:15,800
obviously you know as content marketers you don't want to be building on rented land and

483
00:44:15,800 --> 00:44:20,560
in the case of the perplexity pages tool that is rented land they could probably remove

484
00:44:20,560 --> 00:44:25,440
that tool at any point if they wanted to so I don't know does it get indexed do we know

485
00:44:25,440 --> 00:44:26,440
anything about that?

486
00:44:26,440 --> 00:44:30,840
I've just this second as you asked the question found out whether it gets indexed and the

487
00:44:30,840 --> 00:44:31,840
answer is yes.

488
00:44:31,840 --> 00:44:35,880
Of course it does yeah because guess what now all the listicles and all the how-to's

489
00:44:35,880 --> 00:44:40,760
are going to go live on the perplexity site so thank you dear user for asking great questions

490
00:44:40,760 --> 00:44:45,240
of perplexity and thinking of really useful content that people might want to read about

491
00:44:45,240 --> 00:44:49,680
and then putting it on our site so we can get all the juice no sorry I'm not doing it

492
00:44:49,680 --> 00:44:55,960
personally but I can see it's kind of cool if I can like do it and then remove the page

493
00:44:55,960 --> 00:45:00,640
from perplexity and take all that content and put it on my own site then yeah I might

494
00:45:00,640 --> 00:45:05,680
consider it obviously editing it as well to add my own flavor people we don't use AI

495
00:45:05,680 --> 00:45:11,160
outputs as is for a thousand billion reasons that we've talked about a lot on the podcast

496
00:45:11,160 --> 00:45:16,080
in the past so yeah I think it's quite an interesting little play from perplexity but

497
00:45:16,080 --> 00:45:19,440
personally I wouldn't use it.

498
00:45:19,440 --> 00:45:23,720
There's some other interesting stuff we've had there's some rumors coming out about Siri

499
00:45:23,720 --> 00:45:31,320
Martin there are and this comes off the back of an announcement of a deal between OpenAI

500
00:45:31,320 --> 00:45:40,680
and Apple which has been in the news this week but Siri 2.0 is set to be unveiled at

501
00:45:40,680 --> 00:45:48,400
the upcoming WWDC event which is their worldwide developer conference and it's the big showcase

502
00:45:48,400 --> 00:45:55,200
event for all things Apple and developer things going forward.

503
00:45:55,200 --> 00:46:00,400
What they're saying is that the new Siri will use AI for more granular app control so you'll

504
00:46:00,400 --> 00:46:05,560
be able to actually effectively function calling and tool use that we discussed with Anthropic

505
00:46:05,560 --> 00:46:12,040
and Claude earlier on so developers will be able to connect their apps using APIs to Siri

506
00:46:12,040 --> 00:46:16,400
and enable users to start interacting with them with voice.

507
00:46:16,400 --> 00:46:24,240
AI will analyze user habits to auto enable new Siri commands for Apple apps with plans

508
00:46:24,240 --> 00:46:27,600
to also include third-party apps as well.

509
00:46:27,600 --> 00:46:31,960
These new capabilities will include things like summarizing articles, editing, sharing

510
00:46:31,960 --> 00:46:40,600
photos and managing emails and I can't tell you how often I want to be able to manage

511
00:46:40,600 --> 00:46:45,440
my calendar and my email with a language model.

512
00:46:45,440 --> 00:46:51,280
I am forever wanting to be able to say to chat GPT if I'm out and about add this to

513
00:46:51,280 --> 00:46:57,920
my calendar or give me a task and a to-do list draft this email and stick it in my inbox

514
00:46:57,920 --> 00:47:00,800
so when I'm back in the office I can send it.

515
00:47:00,800 --> 00:47:08,800
If they get this right I think this could be mega for users of Apple devices.

516
00:47:08,800 --> 00:47:13,080
They're talking about the initial launch being in September it will support single commands

517
00:47:13,080 --> 00:47:18,400
with multi-step tasks expected later next year.

518
00:47:18,400 --> 00:47:22,240
I'm excited about this as well I'm like you I mean in some ways I'm excited about it.

519
00:47:22,240 --> 00:47:26,960
I should caveat so when I'm walking the dog I've got so much to do of an average day that

520
00:47:26,960 --> 00:47:29,640
at the moment there's a couple of things I do.

521
00:47:29,640 --> 00:47:34,680
First is I'm constantly dictating emails to Audio Pen which is one of my favorite apps.

522
00:47:34,680 --> 00:47:35,860
Really good app.

523
00:47:35,860 --> 00:47:40,020
Have to have the screen on to be able to dictate into it which basically means I'm walking

524
00:47:40,020 --> 00:47:43,520
around the village with the dog should probably be paying more attention to the dog and I'm

525
00:47:43,520 --> 00:47:48,200
holding my phone near my mouth with the screen on dictating an email and then I have to leave

526
00:47:48,200 --> 00:47:52,660
the screen on so that it can do its transcription and all the other thing that it needs to do.

527
00:47:52,660 --> 00:47:58,160
My ideal use case is I've got my earbud in and I'm just talking to my phone I'm like

528
00:47:58,160 --> 00:48:00,320
right we're going to draft an email now Siri.

529
00:48:00,320 --> 00:48:03,840
Siri's like yeah cool no worries I'll pop open Gmail look what would you like it to

530
00:48:03,840 --> 00:48:07,840
say and then I can dictate it and maybe it automatically cleans the transcription up

531
00:48:07,840 --> 00:48:12,740
and adds the right commas and periods and all this other stuff that it needs in it.

532
00:48:12,740 --> 00:48:17,420
Like I need that in my life ASAP for those of you that commute a lot especially if you're

533
00:48:17,420 --> 00:48:21,440
driving again it could be pros and cons having your concentration on writing an email rather

534
00:48:21,440 --> 00:48:24,680
than driving but certainly that would be really cool.

535
00:48:24,680 --> 00:48:29,440
One thing that you can do which is as close as I can get at the moment is if you use the

536
00:48:29,440 --> 00:48:34,580
current version of GPT-4 in the app version you can actually have a voice conversation

537
00:48:34,580 --> 00:48:38,880
with chat GPT so not the new voice version that's hopefully coming soon although I'm

538
00:48:38,880 --> 00:48:41,200
sure be able to do it.

539
00:48:41,200 --> 00:48:44,760
That I don't have to have the screen open for and I can just have a conversation back

540
00:48:44,760 --> 00:48:52,200
and forth with GPT-4 asking it questions and speaking with it so it doesn't quite it's

541
00:48:52,200 --> 00:48:56,280
not quite the same sort of tool but in theory I could dictate an email it's just every

542
00:48:56,280 --> 00:48:59,920
time you pause it thinks you've finished it wants to give you a response.

543
00:48:59,920 --> 00:49:06,600
Isn't that so frustrating they have made that hyper aggressive in terms of just you can

544
00:49:06,600 --> 00:49:11,280
barely take a pause for breath and it jumps in to respond to you straight away.

545
00:49:11,280 --> 00:49:15,400
That's one of the big things I think when they get 4.0 voice properly rolled out it's

546
00:49:15,400 --> 00:49:18,840
not going to have so I'm quite excited for that but that example that you give there

547
00:49:18,840 --> 00:49:23,420
about having the conversation that's great and when you you know get back to the office

548
00:49:23,420 --> 00:49:28,800
and sit down and open that conversation your email will be drafted and what have you still

549
00:49:28,800 --> 00:49:31,600
not adding it straight into your calendar or it's still not adding it straight into

550
00:49:31,600 --> 00:49:32,600
your Gmail.

551
00:49:32,600 --> 00:49:33,600
Gmail draft.

552
00:49:33,600 --> 00:49:39,800
Yeah and that's the that's the step that's the bit that I'm itching for.

553
00:49:39,800 --> 00:49:47,720
Yeah I agree on the knowing when to speak back and when to pause and stuff I genuinely

554
00:49:47,720 --> 00:49:53,760
don't think well I guess the AI may be able to do it if it can if it can learn enough

555
00:49:53,760 --> 00:49:58,680
about your own personal speech patterns to know how long you typically pause for.

556
00:49:58,680 --> 00:50:03,000
I've got a friend called Andrew sorry Andrew I know you listen to the podcast he loves

557
00:50:03,000 --> 00:50:06,920
to pause even as a human having a conversation with him you don't know if you should speak

558
00:50:06,920 --> 00:50:11,720
or not because he loves to pause for quite a long time and it needs to learn those things

559
00:50:11,720 --> 00:50:15,800
in order to be able to do this properly and dare I say I think it even needs to be able

560
00:50:15,800 --> 00:50:20,240
to have access to your camera because as humans having a conversation it isn't just your voice

561
00:50:20,240 --> 00:50:25,820
that tells me are you finished speaking is it right for me to interject have you got

562
00:50:25,820 --> 00:50:30,260
something to say when I'm speaking I can see that in your body language your facial expressions

563
00:50:30,260 --> 00:50:34,120
there's a loads of other signals that we use in conversation that are really important

564
00:50:34,120 --> 00:50:38,320
to enabling that so I think for this to work really well it's going to have to use your

565
00:50:38,320 --> 00:50:43,280
camera it's going to have to see your face it's going to have to know body language stuff

566
00:50:43,280 --> 00:50:47,360
in order to be able to interpret that properly which of course is not helpful for our drive

567
00:50:47,360 --> 00:50:52,160
the car walk the dog case because we don't want our phones right in front of our faces

568
00:50:52,160 --> 00:50:56,600
but it could be useful when working at the computer like a number of the use cases that

569
00:50:56,600 --> 00:51:02,360
we were excited about on the last episode where having conversations with chat GPT while

570
00:51:02,360 --> 00:51:07,200
I'm doing my work could be helpful but I think it needs to see me to know when I finish speaking

571
00:51:07,200 --> 00:51:12,460
which I have finished speaking now mine which gives me the ideal opportunity to step in

572
00:51:12,460 --> 00:51:17,960
and say actually I said that the deal between Apple and open AI had been announced actually

573
00:51:17,960 --> 00:51:21,960
it is being widely reported but it hasn't been announced yeah and we don't know about

574
00:51:21,960 --> 00:51:25,680
this Siri thing this is a rumor until we get to the actual developer conference and we

575
00:51:25,680 --> 00:51:30,480
see if it gets announced and what have you so yeah want to keep an eye on because hopefully

576
00:51:30,480 --> 00:51:35,400
it gets us closer to the thing that certainly you and I fancy being able to use Martin and

577
00:51:35,400 --> 00:51:41,360
I expect a lot of other people do too right next story is about 11 lamps and their AI

578
00:51:41,360 --> 00:51:47,320
powered text to sound effects so some of you might remember that Martin and I were on tour

579
00:51:47,320 --> 00:51:54,960
in the US doing some consultancy and AI training work also counts as a little plug there this

580
00:51:54,960 --> 00:51:59,320
week's episode is sponsored by Martin and Paul's AI training and consultancy to enable

581
00:51:59,320 --> 00:52:03,360
your business to leverage AI internally as part of your marketing efforts to boost results

582
00:52:03,360 --> 00:52:09,000
and efficiency and just enjoy marketing even more than you do now and advert but while

583
00:52:09,000 --> 00:52:12,840
we were on tour we were playing with this new tool because Martin got early access because

584
00:52:12,840 --> 00:52:18,280
he's been a 11 labs use of quite a while and we basically made it do a bunch of stuff that

585
00:52:18,280 --> 00:52:21,760
it couldn't do and we broke it which is what we normally do but they've now rolled this

586
00:52:21,760 --> 00:52:26,920
out more widely it's got a number of other capabilities that it didn't have when we were

587
00:52:26,920 --> 00:52:33,360
playing with it and it's really quite an interesting tool in terms of being able to create a bunch

588
00:52:33,360 --> 00:52:40,440
of sound effects so as you can imagine the goal here is to enable a bunch of creatives

589
00:52:40,440 --> 00:52:44,200
in a number of industries film television marketing video games social media to be able

590
00:52:44,200 --> 00:52:49,400
to create audio quickly and easily and it's kind of aligned with this concept of we've

591
00:52:49,400 --> 00:52:53,720
got a bunch of tools some of them can do video some of them can do audio some of them can

592
00:52:53,720 --> 00:52:59,000
do text how do you combine all this stuff together to be able to get a an output in

593
00:52:59,000 --> 00:53:06,040
fact I don't know if you saw this Mike there was an example for GPT-4-0 combined with Sora

594
00:53:06,040 --> 00:53:13,560
where someone was able to brief it on creating a video that had a synthetic voice overlay

595
00:53:13,560 --> 00:53:17,720
and then the transcript running along the bottom of the video with music and sound effects

596
00:53:17,720 --> 00:53:18,720
did you see that?

597
00:53:18,720 --> 00:53:20,520
Yeah that that was cool.

598
00:53:20,520 --> 00:53:24,160
Right and you can see how all these tools need to be stacked on top of each other to

599
00:53:24,160 --> 00:53:29,640
get what we would consider the video content we're all used to using that's got music

600
00:53:29,640 --> 00:53:33,280
and voice and sound effect and all that other stuff you need to be able to combine and stack

601
00:53:33,280 --> 00:53:38,160
all those things together to get you know the full output so obviously this 11labs tool

602
00:53:38,160 --> 00:53:41,920
is just for sound effects and sometimes you just need a sound effect but all of this is

603
00:53:41,920 --> 00:53:47,080
also being combined into these tools so that you'll be able to create really interesting

604
00:53:47,080 --> 00:53:51,200
stuff that combines all of the things that you'd expect to see in your average Netflix

605
00:53:51,200 --> 00:53:52,200
show.

606
00:53:52,200 --> 00:53:57,560
It's worth looking at the video that they put out for this with all of the sound effects

607
00:53:57,560 --> 00:54:04,160
created by their own model it can create sound effects up to 22 seconds long and quite

608
00:54:04,160 --> 00:54:12,720
importantly one of the things that they reference is the model has been trained on Shutterstock's

609
00:54:12,720 --> 00:54:17,560
library of licensed audio and ethically sourced sound clips and I think that is quite important

610
00:54:17,560 --> 00:54:21,880
and increasingly the conversations that I'm having with business owners and marketers

611
00:54:21,880 --> 00:54:28,040
is around the copyrights and kind of ethical training of these models.

612
00:54:28,040 --> 00:54:31,960
Yeah I think that's interesting the other thing is I want to be able to automatically

613
00:54:31,960 --> 00:54:38,440
insert sound effects into our podcast Martin a little bit like I think the characters are

614
00:54:38,440 --> 00:54:41,800
called Ira and the Douche from Parks and Rec if you're a Parks and Rec fan you'll know

615
00:54:41,800 --> 00:54:45,120
exactly what I'm talking about and you'll know exactly why we could never include those

616
00:54:45,120 --> 00:54:50,160
sound effects in our podcast because most of them are offensive but that would be quite

617
00:54:50,160 --> 00:54:51,160
funny.

618
00:54:51,160 --> 00:54:57,960
Right and speaking about like crazy content creation the reason we were interested in

619
00:54:57,960 --> 00:55:02,680
the 11 lab story and how all these things combine is because of some interesting news

620
00:55:02,680 --> 00:55:07,440
from a startup called The Simulation that used to be called Fable Studio now we don't

621
00:55:07,440 --> 00:55:12,000
have a huge amount of information on this only a Twitter post and a Forbes article but

622
00:55:12,000 --> 00:55:17,080
tell us about what The Simulation is planning to do with their new tool Showrunner Martin

623
00:55:17,080 --> 00:55:19,040
because this is a bit mind-blowing.

624
00:55:19,040 --> 00:55:23,600
Yeah the website is actually available to go and look at some examples and you can sign

625
00:55:23,600 --> 00:55:28,160
up for early access so what are you signing up for early access to?

626
00:55:28,160 --> 00:55:34,800
Well it's AI generated stories, videos, TV shows from scratch you put in an idea and

627
00:55:34,800 --> 00:55:37,560
it will create it.

628
00:55:37,560 --> 00:55:45,200
They've launched the platform with 10 original shows and the style of them they're kind of

629
00:55:45,200 --> 00:55:51,320
anime or cartoony they're all very cartoon style.

630
00:55:51,320 --> 00:56:00,240
The original version that they previewed in 2023 used South Park and they created a South

631
00:56:00,240 --> 00:56:08,160
Park episode entirely AI generated and it was uncanny how kind of close they got to

632
00:56:08,160 --> 00:56:10,440
creating a real episode.

633
00:56:10,440 --> 00:56:16,480
So yeah now you can generate and watch AI powered TV shows in these virtual worlds it

634
00:56:16,480 --> 00:56:22,200
combines multi-agent simulations with large language models to create this interactive

635
00:56:22,200 --> 00:56:24,600
content.

636
00:56:24,600 --> 00:56:30,120
If you go onto the website you can check them out there's one called Sim San Francisco or

637
00:56:30,120 --> 00:56:38,480
sorry Sim Francisco I think it's called and it's a South Park style animation and you

638
00:56:38,480 --> 00:56:44,640
can see what it's doing and in some instances the animation seems to be really quite basic

639
00:56:44,640 --> 00:56:51,320
where a scene might have three people in it and there was one particular scene I saw where

640
00:56:51,320 --> 00:56:55,200
it was three faces on screen and the only thing that was moving at any time was the

641
00:56:55,200 --> 00:57:00,200
mouth as a different person spoke so relatively simple but then it was stitching different

642
00:57:00,200 --> 00:57:04,400
scenes together and there was some more advanced animation in there but you can see that they've

643
00:57:04,400 --> 00:57:09,200
found quite clever ways to tell stories and reduce the complexity of the animations but

644
00:57:09,200 --> 00:57:14,520
it's only going to get better I mean this is like day one of these projects.

645
00:57:14,520 --> 00:57:21,540
So yeah if you're interested in seeing what an AI generated TV show might look like check

646
00:57:21,540 --> 00:57:28,920
out Showrunner it's available early access you can sign up now but some of the initial

647
00:57:28,920 --> 00:57:33,200
shows that they're creating are available to see on YouTube.

648
00:57:33,200 --> 00:57:39,800
I think it's pretty cool I think you're right I think it it's probably early and so I think

649
00:57:39,800 --> 00:57:43,280
we should all manage our expectations of what that means but like apparently what you're

650
00:57:43,280 --> 00:57:48,240
going to be able to do as a user is create these episodes as Martin described and then

651
00:57:48,240 --> 00:57:52,600
but you can actually go in and edit the scripts and the shots and the voices you can actually

652
00:57:52,600 --> 00:57:55,040
get a bit of fine control over it.

653
00:57:55,040 --> 00:58:01,360
So in terms of putting together sort of you know a lot of a lot of animated shows both

654
00:58:01,360 --> 00:58:08,040
for kids and for adults there's been a focus on the quality of the content dare I say it

655
00:58:08,040 --> 00:58:12,960
even over the quality of the animation right one of the things about South Park is is that

656
00:58:12,960 --> 00:58:21,120
they want the animation to look a bit kind of crappy to use a Cartman word because they

657
00:58:21,120 --> 00:58:24,200
the whole point is this content they want to be able to just produce these episodes

658
00:58:24,200 --> 00:58:28,560
have them be really funny it's the quality of the writing and etc that they want to shine

659
00:58:28,560 --> 00:58:34,680
through not it's not CGI for you know your next Marvel blockbuster but of course with

660
00:58:34,680 --> 00:58:39,920
these tools then maybe you can create those animated shorts and then get some ideas from

661
00:58:39,920 --> 00:58:43,320
prompts but edit the scripts and like bring your vision to life in a way that would have

662
00:58:43,320 --> 00:58:49,960
been really hard without a professional animator in the team and probably quite a large budget.

663
00:58:49,960 --> 00:58:54,440
The idea is going to be that when users are creating episodes one assumes based on these

664
00:58:54,440 --> 00:58:59,920
original show templates to begin with just so they can control and influence and make

665
00:58:59,920 --> 00:59:04,400
it easier to produce the assets I assume because it's probably working from a set of defined

666
00:59:04,400 --> 00:59:09,460
assets is that if you create a great episode that everybody like watches and it goes viral

667
00:59:09,460 --> 00:59:16,280
there'll be some sort of revenue sharing model that you then can get you know both credit

668
00:59:16,280 --> 00:59:20,680
for having produced it but maybe some cash could definitely imagine how if I was like

669
00:59:20,680 --> 00:59:26,920
17 I'd be really excited about making some TV shows and maybe this is how the next big

670
00:59:26,920 --> 00:59:30,600
show producer becomes famous right they create a show on something like this people are like

671
00:59:30,600 --> 00:59:36,680
oh wow ever so witty and what have you and then and then it goes viral.

672
00:59:36,680 --> 00:59:42,480
From a marketing perspective how does this open up creating your own TV shows for brands

673
00:59:42,480 --> 00:59:47,500
right beyond TV shows maybe you can make training videos and other things that are a bit more

674
00:59:47,500 --> 00:59:52,080
interesting that leverage this as a format so it sounds real early like you said mine

675
00:59:52,080 --> 00:59:58,140
but I think the creative people marketers being a subset of people who come up with

676
00:59:58,140 --> 01:00:02,620
really interesting ways to use this can do fantastic and interesting promotion for themselves

677
01:00:02,620 --> 01:00:08,640
for their brands that they work for I've been really interested to see how people let their

678
01:00:08,640 --> 01:00:12,000
creativity loose with this.

679
01:00:12,000 --> 01:00:17,720
Right last couple of stories then so there was a online survey that was conducted recently

680
01:00:17,720 --> 01:00:21,800
and the results were released this week by the Reuters Institute to try and understand

681
01:00:21,800 --> 01:00:27,000
public awareness and the use of generative tools in six countries they ran the survey

682
01:00:27,000 --> 01:00:32,600
in Argentina Denmark France Japan the UK and the USA just to try and get a feel for the

683
01:00:32,600 --> 01:00:38,520
how people feel about generative AI and its impact on different sectors and how many people

684
01:00:38,520 --> 01:00:40,760
have used it and all that good stuff.

685
01:00:40,760 --> 01:00:47,800
I think probably the most interesting things were that most people had heard of AI chat

686
01:00:47,800 --> 01:00:53,460
GPT had been probably known by about most of people so around 50% of people surveyed

687
01:00:53,460 --> 01:01:01,080
had heard of chat GPT but despite having heard of it the frequent use of chat GPT was extremely

688
01:01:01,080 --> 01:01:09,200
rare ranging from just 1% using it daily in Japan to 7% in the USA many users have only

689
01:01:09,200 --> 01:01:13,680
used these tools once or twice indicating that they're not really part of people's

690
01:01:13,680 --> 01:01:18,840
routine internet use as we talked a bit about earlier Martin and the younger people as you

691
01:01:18,840 --> 01:01:23,760
might expect potentially are more likely to use generative AI products regularly with

692
01:01:23,760 --> 01:01:27,720
56% of 18 to 24 year olds having used chat GPT at least once.

693
01:01:27,720 --> 01:01:33,960
I would be interested to know how that increases if you're looking at generative AI baked into

694
01:01:33,960 --> 01:01:39,480
tools like Instagram and TikTok and some of these video generation tools that I know a

695
01:01:39,480 --> 01:01:43,720
lot of people do to make their videos look better when they're sharing them on those

696
01:01:43,720 --> 01:01:48,400
types of platforms but I think what that goes to show is if you are a listener to this podcast

697
01:01:48,400 --> 01:01:53,200
you're trying to make AI use a part of your workflow and you're thinking oh I'm not using

698
01:01:53,200 --> 01:01:57,920
it enough I'm so behind you're probably not behind if you're a listener to a podcast like

699
01:01:57,920 --> 01:02:03,760
this you're probably already in the top 7% in the USA let's say of people who are thinking

700
01:02:03,760 --> 01:02:07,080
about how to adopt these tools into their work.

701
01:02:07,080 --> 01:02:09,120
What did you think about this survey Martin?

702
01:02:09,120 --> 01:02:15,620
I thought it backed up what I would have expected to see that people aren't using these tools

703
01:02:15,620 --> 01:02:21,240
a great deal awareness of them is high I certainly wasn't surprised to see that people had used

704
01:02:21,240 --> 01:02:28,280
them once or twice this goes back to the point we were talking about with GPT 3.5 on chat

705
01:02:28,280 --> 01:02:36,400
GPT just not having enough capability to draw people in and that's why I think 4.0 is going

706
01:02:36,400 --> 01:02:42,120
to be a real game changer particularly when people can start speaking to a chatbot in

707
01:02:42,120 --> 01:02:47,560
their hand that feels like sci-fi that is going to draw people in undoubtedly and you

708
01:02:47,560 --> 01:02:53,280
can imagine the onboarding flow for this new app when they launch it is going to really

709
01:02:53,280 --> 01:02:59,060
focus heavily on the voice capabilities of it and I'd love to have a peek behind the

710
01:02:59,060 --> 01:03:03,760
curtain of chat GPT's product onboarding team right now.

711
01:03:03,760 --> 01:03:08,800
Yeah that's going to be fascinating I want to I want to see people I don't want to see

712
01:03:08,800 --> 01:03:13,480
it in some ways because it's kind of a bit scary but I expect to see people sat in a

713
01:03:13,480 --> 01:03:20,200
pub where somebody knows about this release and this tool and you know the voice capabilities

714
01:03:20,200 --> 01:03:25,200
maybe their friends don't they've got their phone on the table and they're all having

715
01:03:25,200 --> 01:03:29,020
a conversation with it because it's able to speak back to them in real time and remember

716
01:03:29,020 --> 01:03:32,880
that there are different people in the conversation and who they are so it can actually speak

717
01:03:32,880 --> 01:03:36,880
to people by name one assumes based on the demos that we've seen I think that's going

718
01:03:36,880 --> 01:03:40,240
to break people's brains and I think that's going to be one of the first ways that they

719
01:03:40,240 --> 01:03:44,960
really get introduced to it I remember in the village where I live when Dali to came

720
01:03:44,960 --> 01:03:49,360
out trying to explain to people why I could do but not really been able to explain it

721
01:03:49,360 --> 01:03:53,060
very well without showing them so but before I knew it like I'd only shown people like

722
01:03:53,060 --> 01:03:56,480
two images and they were throwing prompts at me across the table like ask it this ask

723
01:03:56,480 --> 01:04:01,720
it this and I think that's going to be the response that we get to this but the difference

724
01:04:01,720 --> 01:04:07,360
is going to be you'll just speak to it yourself and then it will speak back to you I really

725
01:04:07,360 --> 01:04:11,440
think like you say that could be the moment where people are like I want to play with

726
01:04:11,440 --> 01:04:16,720
this more. Right one more story Martin and it's we're into speculative territory right

727
01:04:16,720 --> 01:04:22,880
we're into the simulation and we're into creating your own Netflix shows basically with prompts

728
01:04:22,880 --> 01:04:28,120
and there was a really interesting article this week that we read wasn't there Martin

729
01:04:28,120 --> 01:04:32,920
from the chief of staff at Anthropic tell us about this article.

730
01:04:32,920 --> 01:04:38,120
Yeah Avatar Bulwit which I might have just butchered that so apologies but yeah chief

731
01:04:38,120 --> 01:04:45,600
of staff at Anthropic wrote a really lengthy article in which he says that most jobs will

732
01:04:45,600 --> 01:04:50,680
be obsolete in the next five years and we've heard these predictions before but it's interesting

733
01:04:50,680 --> 01:04:58,240
to see it from somebody who is so close to the to the top within one of these companies

734
01:04:58,240 --> 01:05:06,360
making a frontier model. The article itself goes on to describe this psychology of work

735
01:05:06,360 --> 01:05:13,400
and the psychological impact of widespread job loss and she actually writes a really

736
01:05:13,400 --> 01:05:18,820
interesting article examining what's happened in like the construction industry in different

737
01:05:18,820 --> 01:05:23,480
parts of the world when there's been economic crashes and what it's done for people's sense

738
01:05:23,480 --> 01:05:28,200
of worth and things like that but she goes on to basically say that we need to have a

739
01:05:28,200 --> 01:05:36,940
real conversation around this but that part is interesting I think the thing that just

740
01:05:36,940 --> 01:05:42,360
triggered us to feature this in the podcast was just that somebody is coming out and saying

741
01:05:42,360 --> 01:05:48,860
this she's a young woman a young professional a whole career ahead of herself and from what

742
01:05:48,860 --> 01:05:55,440
she's seeing day to day right now she's looking at it going knowledge workers like me will

743
01:05:55,440 --> 01:06:01,560
be unnecessary within three to five years was their timeline I think the opening line

744
01:06:01,560 --> 01:06:06,820
of the the article actually says three years never mind five years. Yeah I think it's definitely

745
01:06:06,820 --> 01:06:12,600
worth quoting that this article's title is my last five years of work pretty compelling

746
01:06:12,600 --> 01:06:18,800
let's be honest and then the first paragraph is I am 25 these next three years might be

747
01:06:18,800 --> 01:06:23,720
the last few years that I work I'm not ill nor am I becoming a stay at home mom nor have

748
01:06:23,720 --> 01:06:28,920
I been so financially fortunate to be on the brink of voluntary retirement I stand at the

749
01:06:28,920 --> 01:06:33,120
edge of a technological development that seems likely should it arrive to end employment

750
01:06:33,120 --> 01:06:40,600
as I know it crumbs that is a heck of an intro paragraph to any piece that you might read

751
01:06:40,600 --> 01:06:44,800
it's beautifully written honestly I think is a really good article lots of interesting

752
01:06:44,800 --> 01:06:53,560
thoughts if you're interested on the impact of AI on work and therefore the impact of

753
01:06:53,560 --> 01:07:00,240
disrupting work as we know it to us as humans is definitely worth a read yeah and do you

754
01:07:00,240 --> 01:07:07,040
do you agree broadly speaking given everything that we have spoken about my experience at

755
01:07:07,040 --> 01:07:13,520
the top of the show talking about data analysis these frustrations that we have where you

756
01:07:13,520 --> 01:07:23,760
know 90% there is is not there like 90% right is is not right and then do do you think within

757
01:07:23,760 --> 01:07:30,720
that time frame we'll have got to the point where knowledge workers will be stripped from

758
01:07:30,720 --> 01:07:35,240
the economy at such scale because I just I don't see it from where I sit do you know

759
01:07:35,240 --> 01:07:42,600
what it is the thing I find hardest about this without any doubt is two opposing ideas

760
01:07:42,600 --> 01:07:48,760
that I'm having to hold in my mind at the same time the first is that this person works

761
01:07:48,760 --> 01:07:57,560
at anthropic and the commercial value of tools that that take on the productivity of all

762
01:07:57,560 --> 01:08:03,160
of the knowledge workers in the world but we still live in some sort of economic system

763
01:08:03,160 --> 01:08:07,360
with money like we live in now means that those tools doing all that work will be worth

764
01:08:07,360 --> 01:08:12,680
an absolute fortune so there is a massive commercial incentive for the people inside

765
01:08:12,680 --> 01:08:18,680
of these organizations to be saying LNMs and AI is going to do all the productive work

766
01:08:18,680 --> 01:08:24,360
and you know we should get ready for it and you know with Sam Altman has probably not

767
01:08:24,360 --> 01:08:28,840
explored this particular topic at this level of detail and potentially in this level of

768
01:08:28,840 --> 01:08:33,360
eloquence as this wonderful article does but it's not dissimilar to some of the things

769
01:08:33,360 --> 01:08:37,880
that we hear coming out from the likes of open AI but at the same time we've got Yana

770
01:08:37,880 --> 01:08:41,760
Kulnove we've mentioned number of times on the podcast who's head of AI at Meta going

771
01:08:41,760 --> 01:08:48,060
our AIs are no smarter than a house cat so unless house cats are going to somehow rise

772
01:08:48,060 --> 01:08:52,760
up and do all of the knowledge work across the world then this is all absolute baloney

773
01:08:52,760 --> 01:08:57,600
right and I think that's the challenge is she's on the inside of a system that we're

774
01:08:57,600 --> 01:09:01,840
not that sees things that are potentially three six nine twelve months ahead of what

775
01:09:01,840 --> 01:09:07,400
we see so what is she seeing that's giving her this confidence maybe stuff we don't know

776
01:09:07,400 --> 01:09:12,560
about that if we saw it we'd feel the same or are they all incentivized to talk this

777
01:09:12,560 --> 01:09:17,480
up massively because it's just going to inflate the value of their businesses and I just don't

778
01:09:17,480 --> 01:09:24,920
know mine but to your point I think for me until we eliminate some of these errors we've

779
01:09:24,920 --> 01:09:28,080
got a big problem because I just see a bunch of people are just not going to trust them

780
01:09:28,080 --> 01:09:32,440
before we even get into heavily regulated industries right financial health care what

781
01:09:32,440 --> 01:09:35,760
have you where those errors are just not going to be tolerated and somebody's going to be

782
01:09:35,760 --> 01:09:40,400
held accountable for it there's an AI makes an error and a patient gets given a wrong

783
01:09:40,400 --> 01:09:45,720
medication and they get sick and they die people are going to be held accountable for

784
01:09:45,720 --> 01:09:48,920
that somebody's going to go to jail right and people are just not going to allow that

785
01:09:48,920 --> 01:09:54,180
to happen so we have to get those out does GPT-5 with some sort of logic built into it

786
01:09:54,180 --> 01:09:56,920
to make sure it doesn't make any mistakes anymore fix that I don't know right we'll

787
01:09:56,920 --> 01:10:00,640
have to see so I think that part's going to be absolutely critical and then the other

788
01:10:00,640 --> 01:10:05,160
things change management we've talked about on the podcast today Martin that even people

789
01:10:05,160 --> 01:10:09,480
who are heavily exposed to this people we know some of them are people that we train

790
01:10:09,480 --> 01:10:15,440
and support need to change their behavior to have chat GPT open all day every day to

791
01:10:15,440 --> 01:10:19,680
even think oh maybe I'll just ask chat GPT instead of emailing that colleague because

792
01:10:19,680 --> 01:10:24,800
maybe I'll get a better answer out of chat GPT and they don't so we've got that whole

793
01:10:24,800 --> 01:10:30,440
human change element that we've got to manage as well so and I think that's slow I think

794
01:10:30,440 --> 01:10:34,320
technology change might be fast but human change and behavior change appears to be quite

795
01:10:34,320 --> 01:10:40,040
slow at least in my limited experience on this planet so five years don't know feels

796
01:10:40,040 --> 01:10:45,560
quite long way away ten years I would say probably seems realistic to me five I think

797
01:10:45,560 --> 01:10:50,160
it's the people part that will slow it down because I think we'll resist yeah there'll

798
01:10:50,160 --> 01:10:55,900
be it wouldn't surprise me if there were with the likes of function calling and tool use

799
01:10:55,900 --> 01:11:01,120
and people building products that effectively you can imagine entire products come into

800
01:11:01,120 --> 01:11:08,760
the market that are your AI CFO right and then QuickBooks and these companies go you

801
01:11:08,760 --> 01:11:13,840
don't need that that job role anymore because we've got an AI that will do it for you but

802
01:11:13,840 --> 01:11:20,760
I think as you say it's the the people change the is the chief exec going yeah I trust that

803
01:11:20,760 --> 01:11:26,960
feature of that software to do that job and there's going to be resistance from the people

804
01:11:26,960 --> 01:11:32,500
whose jobs it's going to take right absolutely and I think if we assume that five years ten

805
01:11:32,500 --> 01:11:41,560
years is the right time frame right not 30 years we've got a lot of reconstruction of

806
01:11:41,560 --> 01:11:47,240
the economy to do in order to not have this be a complete car crash and that bit does

807
01:11:47,240 --> 01:11:53,140
worry me and I think one of the sort of more interesting parts of of avatars article is

808
01:11:53,140 --> 01:11:58,480
really trying to explore some of the things around what work means to people in terms

809
01:11:58,480 --> 01:12:05,000
of giving them purpose making them feel like you know they work hard and they get paid

810
01:12:05,000 --> 01:12:11,600
well and then they can feed their family or go on holiday and just help the human brain

811
01:12:11,600 --> 01:12:17,320
is wired with its reward structures it's incentivization structures and also some of its more negative

812
01:12:17,320 --> 01:12:21,880
structures like being ashamed when you don't have a job and stuff like that like it's

813
01:12:21,880 --> 01:12:27,280
going to be extremely complicated and it's not going to all happen at once so the people

814
01:12:27,280 --> 01:12:32,440
who lose their jobs first there probably won't be any safety network outside of the safety

815
01:12:32,440 --> 01:12:38,440
networks that we've got already and how do you feel going to the job centre which is

816
01:12:38,440 --> 01:12:43,880
what we have the equivalent of in the UK and saying you need to help retrain me I need

817
01:12:43,880 --> 01:12:49,960
to find a job because the job I used to get paid 80k British pounds a year to do doesn't

818
01:12:49,960 --> 01:12:55,040
exist anymore so I don't really know what to do with myself and I need to get another

819
01:12:55,040 --> 01:13:01,500
job and what do you do when you're when your skills are redundant I just it's just it's

820
01:13:01,500 --> 01:13:07,360
going to be if we get to that point it's going to be it's going to be difficult I think it's

821
01:13:07,360 --> 01:13:10,640
going to be difficult I haven't explained it very well but hopefully dear listeners

822
01:13:10,640 --> 01:13:15,960
you can sort of understand some of the sort of stressful trepidation that comes with trying

823
01:13:15,960 --> 01:13:20,120
to imagine what that looks like and goodness gracious Martin I'm glad I don't have to

824
01:13:20,120 --> 01:13:21,120
sort it out.

825
01:13:21,120 --> 01:13:29,500
Yeah there's people paid far more money than either of us trying to fix this problem.

826
01:13:29,500 --> 01:13:37,880
So basic income a nice big robot tax pay per token you know every token that you use there's

827
01:13:37,880 --> 01:13:41,840
a nice little slice to the government and the government can just cascade that down

828
01:13:41,840 --> 01:13:47,020
to to all of us and we can chase butterflies all day.

829
01:13:47,020 --> 01:13:52,800
Yeah I like Sam Altman's idea of like universal compute income or whatever where basically

830
01:13:52,800 --> 01:13:57,000
everybody gets a share of compute and you can use it to create your own agents to do

831
01:13:57,000 --> 01:14:02,560
stuff for you or you can sell it to other people or team up with a bunch of other people

832
01:14:02,560 --> 01:14:08,320
and all put your compute together and you know compute becomes the currency of the world

833
01:14:08,320 --> 01:14:13,320
given the number of investments he's making in energy companies and chip companies though

834
01:14:13,320 --> 01:14:16,040
you'd have to wonder if there's a bit of an incentive there as well.

835
01:14:16,040 --> 01:14:17,960
He would say that wouldn't he.

836
01:14:17,960 --> 01:14:21,280
And this is why it's so hard I really want to trust someone Martin I just want to trust

837
01:14:21,280 --> 01:14:24,680
someone to just give me some really good advice give me a feel for where things are going

838
01:14:24,680 --> 01:14:29,840
but I think the it's a mixture of nobody really knows and the people who probably know

839
01:14:29,840 --> 01:14:33,920
the most are incentivised to say certain things and not other things so how are we supposed

840
01:14:33,920 --> 01:14:34,920
to know.

841
01:14:34,920 --> 01:14:37,600
What a way to close out the podcast.

842
01:14:37,600 --> 01:14:38,960
And on that cheery note.

843
01:14:38,960 --> 01:14:44,600
Yeah sorry everybody came in they were like okay AI sucks first 20 minutes next 40 minutes

844
01:14:44,600 --> 01:14:49,240
oh we can do some pretty interesting and cool stuff last 10 minutes yeah amazing I'm going

845
01:14:49,240 --> 01:14:53,680
to create my own TV show what we're going to do when nobody's got jobs anymore.

846
01:14:53,680 --> 01:14:58,160
So there you are we'd like to take you on a journey on this podcast we'll be back next

847
01:14:58,160 --> 01:15:04,760
week to inspire scare and depress you in equal measure probably hopefully not too much.

848
01:15:04,760 --> 01:15:09,840
If you still want to listen to this podcast after this episode ringing around in your

849
01:15:09,840 --> 01:15:15,240
ears and your brains over the next week or two then please subscribe and do share this

850
01:15:15,240 --> 01:15:19,640
with your friends because you know it hasn't been all positive this episode but these are

851
01:15:19,640 --> 01:15:24,040
important things to talk about we do want to get into the other sides of AI every now

852
01:15:24,040 --> 01:15:26,760
and again don't we Martin.

853
01:15:26,760 --> 01:15:30,440
With that in mind I'm going to leave you to your Sunday I'm going to leave you dear listener

854
01:15:30,440 --> 01:15:33,860
to chew on some of the things you've heard about today and we will be back with you in

855
01:15:33,860 --> 01:15:35,000
a couple of weeks.

856
01:15:35,000 --> 01:15:36,680
Thanks for your time Martin.

857
01:15:36,680 --> 01:15:42,280
Thank you for listening to artificially intelligent marketing to stay on top of the latest trends

858
01:15:42,280 --> 01:15:46,200
tips and tools in the world of marketing AI.

859
01:15:46,200 --> 01:15:50,240
Be sure to subscribe we look forward to seeing you again next week.