← All Webinars | L.A.B.S. #8

Level: Beginner🐣

Equip yourself with practical tools, strategies, and real-world examples to boost your AI skills and effectiveness.

+ 

1
00:00:15.285 --> 00:00:15.505
Hi.

2
00:00:15.505 --> 00:00:18.385
Welcome everyone. We'll get started in just a few minutes

3
00:00:18.535 --> 00:00:21.585
when I give people time to join us and log in.

4
00:00:23.705 --> 00:00:27.515
Welcome, Andrea. Hi, Amanda. Lovely to see you. You too.

5
00:00:32.535 --> 00:00:34.025
Welcome. Welcome everyone.

6
00:00:40.165 --> 00:00:41.505
I'm gonna give everyone a few minutes

7
00:00:41.565 --> 00:00:43.065
to join coming in from lunch.

8
00:01:06.605 --> 00:01:08.745
We flowing in here. Welcome everyone.

9
00:01:09.015 --> 00:01:10.665
Just a few minutes before we get started.

10
00:01:32.865 --> 00:01:35.565
One more moment and then we'll kick things off.

11
00:01:45.445 --> 00:01:47.445
A lot of familiar faces in the chat.

12
00:01:47.445 --> 00:01:48.805
Thanks for joining us today.

13
00:01:53.675 --> 00:01:55.255
All right, well we are at two

14
00:01:55.255 --> 00:01:56.495
after, so we're gonna go ahead and get started.

15
00:01:56.995 --> 00:01:59.135
Um, thank you everyone for joining us today.

16
00:01:59.235 --> 00:02:01.815
My name is Amanda Crowley. I'm the director of marketing.

17
00:02:01.885 --> 00:02:04.415
I'll be your host and moderator today for this session.

18
00:02:05.235 --> 00:02:07.175
Um, so thank you again for joining us.

19
00:02:07.175 --> 00:02:09.535
Many of you are returning, which I appreciate.

20
00:02:09.795 --> 00:02:12.575
Uh, if you are new here, welcome to ITS Summer Demo Days

21
00:02:13.195 --> 00:02:14.255
AI and Practice series.

22
00:02:14.285 --> 00:02:15.375
This is part three.

23
00:02:15.715 --> 00:02:17.495
Before we get started with Andrea,

24
00:02:17.695 --> 00:02:20.215
I just wanna give you some housekeeping things to go over.

25
00:02:20.985 --> 00:02:22.455
First, we're gonna be using the q

26
00:02:22.455 --> 00:02:24.655
and a that's at the bottom of your Zoom.

27
00:02:25.075 --> 00:02:27.695
Uh, if you have a question, go ahead and put it in there

28
00:02:27.695 --> 00:02:28.975
and we'll answer it live time.

29
00:02:29.395 --> 00:02:32.255
If we happen to get too many, um, that's okay.

30
00:02:32.255 --> 00:02:33.455
We'll answer them after the webinar.

31
00:02:33.715 --> 00:02:35.655
The webinar itself is 45 minutes long,

32
00:02:36.155 --> 00:02:38.295
so this session will be recorded.

33
00:02:38.355 --> 00:02:39.655
If you can't stay the whole time

34
00:02:39.795 --> 00:02:41.895
or you wanna revisit it, we will go ahead

35
00:02:41.895 --> 00:02:43.015
and send it out via email.

36
00:02:43.635 --> 00:02:45.975
And then lastly, when we are done today,

37
00:02:45.995 --> 00:02:47.135
we will send you a survey.

38
00:02:47.515 --> 00:02:49.535
It comes up right when you click out of Zoom.

39
00:02:49.715 --> 00:02:51.415
If you could take a second to answer it,

40
00:02:51.415 --> 00:02:52.695
it's about five questions long.

41
00:02:53.515 --> 00:02:56.135
So thank you. Thank you again for being here.

42
00:02:56.555 --> 00:02:57.815
Um, we are with Andrea.

43
00:02:57.915 --> 00:03:00.525
She is gonna be talking about prompt engineering today,

44
00:03:00.525 --> 00:03:02.405
which is such a fun topic.

45
00:03:02.665 --> 00:03:04.565
Uh, she is our senior product manager

46
00:03:05.105 --> 00:03:07.245
and we also wanted to give a special sh shout

47
00:03:07.245 --> 00:03:08.485
out to Emily Bank.

48
00:03:08.575 --> 00:03:11.765
She's our technical writer who did a lot of, um, time

49
00:03:11.825 --> 00:03:13.965
and effort to help us prepare for this session today.

50
00:03:14.425 --> 00:03:16.525
So with that, I'll turn it over to you, Andrea.

51
00:03:16.935 --> 00:03:19.365
Thank you Amanda. So a little bit about me.

52
00:03:19.485 --> 00:03:21.605
I joined ITS about seven years ago

53
00:03:21.865 --> 00:03:24.445
and I'm on the product management team for item workshop.

54
00:03:24.445 --> 00:03:26.085
That's our item banking platform.

55
00:03:26.705 --> 00:03:28.205
And at the start of this year,

56
00:03:28.425 --> 00:03:31.205
my team rolled out our first Spark AI feature in item

57
00:03:31.485 --> 00:03:34.125
workshop, which was automatic item generation using

58
00:03:34.125 --> 00:03:35.205
generative ai.

59
00:03:36.025 --> 00:03:39.005
And my team's currently busy developing our next round

60
00:03:39.005 --> 00:03:40.805
of features that are gonna harness the power

61
00:03:40.865 --> 00:03:41.925
of generative AI

62
00:03:41.985 --> 00:03:44.805
and machine learning to support analysis tasks

63
00:03:44.905 --> 00:03:46.925
and content generation and item workshop.

64
00:03:47.785 --> 00:03:50.365
But when we started, my team had to jump in

65
00:03:50.425 --> 00:03:53.245
and learn about generative AI and how to write prompts.

66
00:03:53.385 --> 00:03:55.565
So I'm really excited to share with you

67
00:03:55.565 --> 00:03:56.885
what we've learned today.

68
00:03:58.315 --> 00:04:00.765
Wonderful. So to get started, Andrea,

69
00:04:00.835 --> 00:04:02.645
what is a model? What is a provider?

70
00:04:03.455 --> 00:04:05.165
Let's start with artificial intelligence.

71
00:04:05.165 --> 00:04:07.725
Let's go back to the beginning. So what is ai?

72
00:04:08.225 --> 00:04:11.765
So AI is just a set of technologies that enables computers

73
00:04:11.785 --> 00:04:13.845
to emulate human intelligence

74
00:04:13.985 --> 00:04:16.005
and perform tasks that require learning

75
00:04:16.185 --> 00:04:17.845
and reasoning and problem solving.

76
00:04:18.465 --> 00:04:21.485
So then generative AI is just a type of AI

77
00:04:21.675 --> 00:04:23.765
that uses large language models

78
00:04:24.065 --> 00:04:25.645
to generate and analyze content.

79
00:04:26.785 --> 00:04:28.605
So what is a large language model?

80
00:04:29.155 --> 00:04:30.525
Well, a large language model

81
00:04:30.825 --> 00:04:34.445
or an LLM, it's just a generative AI program that's capable

82
00:04:34.625 --> 00:04:37.365
of understanding, analyzing text and images.

83
00:04:38.145 --> 00:04:41.205
And these models, they're trained on large sources of data

84
00:04:41.625 --> 00:04:44.045
and that data is often scraped from the internet.

85
00:04:44.305 --> 00:04:48.645
So you may already be familiar with an LLM, um, for example,

86
00:04:49.565 --> 00:04:53.485
GPT, you may have used chat GBT, well GBT four,

87
00:04:53.625 --> 00:04:54.925
that's an example of a model.

88
00:04:55.835 --> 00:05:00.585
Yeah. Okay. Um, and what about providers?

89
00:05:01.445 --> 00:05:04.465
So the companies that own these models, they're known

90
00:05:04.465 --> 00:05:05.585
as AI providers.

91
00:05:05.845 --> 00:05:10.105
So open AI is the AI provider behind G BT four,

92
00:05:10.725 --> 00:05:14.385
and then Google is the AI provider behind the Model Gemini.

93
00:05:14.845 --> 00:05:17.745
So we're gonna see a list of providers today in the demo.

94
00:05:18.815 --> 00:05:20.025
Okay. All right, got it.

95
00:05:20.405 --> 00:05:23.225
So when you log into a provider

96
00:05:23.365 --> 00:05:27.505
or a model, you begin prompting or or talking to that model.

97
00:05:27.775 --> 00:05:29.745
What does prompt engineering exactly?

98
00:05:30.815 --> 00:05:33.025
Yeah, so prompt engineering is a lot of fun,

99
00:05:33.365 --> 00:05:37.185
but I wanna first start by remembering back to late 2022

100
00:05:37.695 --> 00:05:40.225
when OpenAI launched chat, GBT

101
00:05:40.835 --> 00:05:44.345
generative AI pretty much overnight put the world a buzz

102
00:05:45.485 --> 00:05:47.225
and OpenAI already had other

103
00:05:47.225 --> 00:05:49.065
models, they already had them out.

104
00:05:49.065 --> 00:05:50.265
But what was so brilliant

105
00:05:50.485 --> 00:05:55.265
and addictive about chat GBT was that every man, all of us,

106
00:05:55.725 --> 00:05:57.845
we suddenly could chat with ai.

107
00:05:58.265 --> 00:06:00.325
And it felt like this very democratizing moment

108
00:06:00.995 --> 00:06:05.045
because I'm not a developer and I'm not a data modeler,

109
00:06:05.185 --> 00:06:06.365
but I chat

110
00:06:06.365 --> 00:06:08.245
with my colleagues on the Microsoft

111
00:06:08.245 --> 00:06:09.485
teams throughout my workday.

112
00:06:10.025 --> 00:06:11.805
We all know how to chat and now I could chat

113
00:06:11.805 --> 00:06:13.045
with that AI model.

114
00:06:13.665 --> 00:06:17.085
So essentially prompt engineering is just that.

115
00:06:17.395 --> 00:06:20.685
It's simply chatting with your new colleague, the LLM.

116
00:06:21.545 --> 00:06:24.565
So the prompt itself is just a request, it's just your text

117
00:06:24.565 --> 00:06:26.085
that you send to the generative model

118
00:06:26.465 --> 00:06:28.325
and then the generative model is going

119
00:06:28.325 --> 00:06:30.125
to send you back a response.

120
00:06:31.225 --> 00:06:34.125
So you may be a prompt engineering newbie, um,

121
00:06:34.245 --> 00:06:37.765
coming into this webinar, but the skills you use every day

122
00:06:37.865 --> 00:06:39.805
to write your text, write your emails,

123
00:06:40.065 --> 00:06:41.845
and to help train up your teammates,

124
00:06:42.195 --> 00:06:44.925
they're gonna make you a fantastic prompt engineer.

125
00:06:46.075 --> 00:06:49.165
That AI model is just your highly logical teammate

126
00:06:49.165 --> 00:06:52.165
who needs your support and your coaching to be successful.

127
00:06:53.385 --> 00:06:55.515
Yeah, so I was sharing with Andrea

128
00:06:55.615 --> 00:06:57.235
before the session got started.

129
00:06:57.695 --> 00:06:59.475
She was the reason I really started playing

130
00:06:59.505 --> 00:07:02.995
with tools like Chachi, bt I was so intimidated.

131
00:07:03.215 --> 00:07:05.995
I'm not, you know, a computer engineer or anything.

132
00:07:06.135 --> 00:07:08.395
And Andrea said, oh no, like this is built

133
00:07:08.395 --> 00:07:10.315
for people like you that I'm not.

134
00:07:10.535 --> 00:07:13.395
Um, well are technically challenged for better words

135
00:07:13.975 --> 00:07:15.235
and go in and try it.

136
00:07:15.495 --> 00:07:17.155
And so when I first started trying it,

137
00:07:17.315 --> 00:07:20.995
I was getting responses that were just not useful for me.

138
00:07:21.455 --> 00:07:23.675
So is there anything that you can,

139
00:07:23.675 --> 00:07:25.915
advice you can give us about improving the quality

140
00:07:25.915 --> 00:07:27.995
of responses from the AI model?

141
00:07:28.865 --> 00:07:31.915
Yeah, I'm gonna give you five tips today. Okay.

142
00:07:31.975 --> 00:07:33.315
Um, so we're gonna go through five tips

143
00:07:33.315 --> 00:07:35.075
to keep in mind when writing those prompts.

144
00:07:35.415 --> 00:07:36.955
And Amanda, if there's time at the end,

145
00:07:37.075 --> 00:07:39.155
I have a couple bonus tips that I'd love to share.

146
00:07:39.545 --> 00:07:41.355
Sure. So there's a couple things

147
00:07:41.495 --> 00:07:42.715
I'd like to share up front though.

148
00:07:43.335 --> 00:07:46.355
So the first is don't give up if you don't get

149
00:07:46.435 --> 00:07:47.475
a good response initially.

150
00:07:48.465 --> 00:07:51.275
When you send an email to your colleague, um, trying

151
00:07:51.275 --> 00:07:52.595
to elicit information from them

152
00:07:52.895 --> 00:07:55.315
and they, their response fails to answer your questions,

153
00:07:55.655 --> 00:07:57.875
you don't give up, you follow up,

154
00:07:57.935 --> 00:08:00.475
you send an additional email, giving them some more details

155
00:08:01.055 --> 00:08:02.795
and making sure your questions are clear

156
00:08:02.815 --> 00:08:04.715
and concise so that they can respond to you.

157
00:08:05.615 --> 00:08:07.715
Do the same thing with the generative ai.

158
00:08:07.975 --> 00:08:09.955
If you don't get what you're looking for the first time,

159
00:08:10.605 --> 00:08:12.115
apply some of the tips from today

160
00:08:12.295 --> 00:08:14.355
and keep trying, keep trying again.

161
00:08:15.295 --> 00:08:17.475
The second thing I'd like to add up front is

162
00:08:17.475 --> 00:08:20.715
that the quality of the responses can differ across models.

163
00:08:21.575 --> 00:08:24.915
So for example, GBT four Omni, which was

164
00:08:25.455 --> 00:08:29.195
the most recently released, um, model from OpenAI,

165
00:08:29.385 --> 00:08:31.955
it's faster, it's smarter, it's bigger,

166
00:08:32.065 --> 00:08:34.355
it's better than the earlier GPT models like

167
00:08:34.355 --> 00:08:35.635
Da Vinci and GPT-3.

168
00:08:36.335 --> 00:08:37.835
So the models are changing

169
00:08:38.295 --> 00:08:40.355
and they do have different, different abilities

170
00:08:41.095 --> 00:08:43.835
and you're gonna wanna try several models out to find

171
00:08:43.835 --> 00:08:45.435
what works best for your use case.

172
00:08:46.055 --> 00:08:47.715
Um, different models have different strengths.

173
00:08:47.815 --> 00:08:50.715
Amanda, if time allows at the end of this webinar, I'd love

174
00:08:50.715 --> 00:08:52.995
to cover some of the basics of model differences.

175
00:08:57.255 --> 00:08:59.265
Okay. Yeah, we can definitely do that.

176
00:08:59.485 --> 00:09:03.865
Um, at ITS, we have a Spark AI playground,

177
00:09:04.245 --> 00:09:05.665
so it's something that some

178
00:09:05.665 --> 00:09:07.425
of our internal teams have been using.

179
00:09:07.725 --> 00:09:09.105
Can you tell me a little bit about that?

180
00:09:09.885 --> 00:09:11.025
I'd love to. So we're actually

181
00:09:11.025 --> 00:09:12.065
gonna use the playground today.

182
00:09:12.575 --> 00:09:15.145
Okay. Um, and the Spark AI playground,

183
00:09:15.485 --> 00:09:17.665
it was originally created as an internal tool

184
00:09:17.665 --> 00:09:18.825
for ITS employees

185
00:09:19.205 --> 00:09:20.665
and it was created by our director

186
00:09:20.725 --> 00:09:24.505
of Innovative Technologies, Chris CLN and our IDC team.

187
00:09:25.125 --> 00:09:27.305
And this was really an internal tool initially

188
00:09:27.405 --> 00:09:30.545
and it was to help ITS employees learn about prompt

189
00:09:30.865 --> 00:09:32.625
engineering and to try out their ideas.

190
00:09:33.415 --> 00:09:36.225
It's been incredibly useful tools to all of us internally

191
00:09:36.925 --> 00:09:39.225
and here at ITS we're dedicated

192
00:09:39.485 --> 00:09:41.305
to innovating with our clients.

193
00:09:41.845 --> 00:09:44.345
So we decided to extend access

194
00:09:44.365 --> 00:09:46.665
to our Spark AI playground to our clients.

195
00:09:47.445 --> 00:09:49.345
So I'm gonna share my screen in just a moment

196
00:09:49.365 --> 00:09:51.025
and I'll give you guys a tour of the playground.

197
00:09:52.025 --> 00:09:52.725
Sounds great.

198
00:10:00.315 --> 00:10:01.695
Amanda, can you see my screen?

199
00:10:02.155 --> 00:10:05.295
Yep, I sure can. Great. Okay.

200
00:10:05.915 --> 00:10:08.455
So the first thing I'd like to point out in the Spark AI

201
00:10:08.455 --> 00:10:10.735
playground is that you can use it to generate text

202
00:10:11.115 --> 00:10:12.255
or to generate images.

203
00:10:12.745 --> 00:10:14.615
Today we're gonna be generating text.

204
00:10:15.525 --> 00:10:17.295
Next thing I'd like to point out is this

205
00:10:17.355 --> 00:10:18.775
prompt text box up here.

206
00:10:18.775 --> 00:10:21.015
This is where we're going to be entering our prompts today.

207
00:10:21.635 --> 00:10:23.335
And once we're done entering our prompts,

208
00:10:23.335 --> 00:10:24.775
we're gonna click the submit button.

209
00:10:25.725 --> 00:10:28.815
This response area is where the AI model is going

210
00:10:28.815 --> 00:10:30.575
to share their responses back to me.

211
00:10:31.435 --> 00:10:33.655
So for the first model we're gonna look at today,

212
00:10:33.915 --> 00:10:37.045
I'm gonna select OpenAI

213
00:10:38.955 --> 00:10:43.705
and I'm gonna select that new, okay,

214
00:10:44.055 --> 00:10:45.105
technical difficulty.

215
00:10:45.245 --> 00:10:46.785
I'm gonna have to sign out, I'm gonna have

216
00:10:46.785 --> 00:10:50.765
to refresh my page and

217
00:10:50.765 --> 00:10:52.005
potentially sign back in.

218
00:11:05.845 --> 00:11:09.035
There we go. I think I just had had this waiting a little

219
00:11:09.035 --> 00:11:11.715
too long on my screen that I timed out.

220
00:11:12.575 --> 00:11:16.035
So I'm gonna go ahead and select OpenAI as my provider.

221
00:11:16.175 --> 00:11:18.475
And then from the model list of OpenAI models,

222
00:11:18.695 --> 00:11:20.755
I'm gonna select GPT-4 Omni.

223
00:11:21.495 --> 00:11:25.715
Now when I send my re response GPT-4, I'm send my prompt.

224
00:11:26.035 --> 00:11:27.795
GPT-4 is going to share its response

225
00:11:27.915 --> 00:11:29.235
with me in this text box here.

226
00:11:29.535 --> 00:11:32.795
But I wanna compare the performance of G PT four

227
00:11:32.795 --> 00:11:35.195
to another model to see how they respond

228
00:11:35.195 --> 00:11:36.395
to my, my single prompt.

229
00:11:36.855 --> 00:11:38.795
So I'm gonna click add compare here,

230
00:11:39.575 --> 00:11:42.035
and in this response section I'm gonna choose another model.

231
00:11:42.565 --> 00:11:44.235
Let's go with Google Gemini.

232
00:11:46.125 --> 00:11:48.465
And Google Gemini is gonna send its response

233
00:11:48.465 --> 00:11:50.505
to me in this text box here at the bottom.

234
00:11:50.925 --> 00:11:53.965
Now I could start adding additional models as many

235
00:11:53.965 --> 00:11:55.645
as I want across all 10

236
00:11:55.705 --> 00:11:58.245
or 11 of those providers that you saw on the provider list

237
00:11:58.585 --> 00:12:00.685
to compare how they all respond to my prompts.

238
00:12:00.905 --> 00:12:02.605
But I'm gonna keep it pretty simple right now

239
00:12:02.605 --> 00:12:03.965
for the purposes of this demo.

240
00:12:05.065 --> 00:12:07.045
So let's start with prompt number one.

241
00:12:07.345 --> 00:12:09.365
I'm gonna copy and paste it into this prompt

242
00:12:09.595 --> 00:12:10.605
text box right here.

243
00:12:18.675 --> 00:12:22.565
Okay. So my first tip is to assign

244
00:12:23.495 --> 00:12:24.935
a persona to the ai.

245
00:12:25.915 --> 00:12:28.535
So when I'm added to a new work project, one

246
00:12:28.535 --> 00:12:31.015
of my first questions is, what is my role on this team?

247
00:12:31.795 --> 00:12:35.655
And the answer to that question changes the work I do

248
00:12:35.955 --> 00:12:37.375
and what I share back with the team.

249
00:12:38.035 --> 00:12:41.175
So for example, if a colleague emails me a draft

250
00:12:41.355 --> 00:12:45.055
of an ITS blog post, I'm gonna wanna know is my role

251
00:12:45.155 --> 00:12:47.455
as editor, am I looking for spelling and grammar?

252
00:12:48.035 --> 00:12:50.815
Or is my role a writer role where they're expecting me

253
00:12:50.815 --> 00:12:53.295
to add content and rewrite this blog post?

254
00:12:53.955 --> 00:12:58.295
So treat the spark ai, uh, treat the AI model the same way.

255
00:12:58.365 --> 00:13:00.375
Tell it what you're expecting, what is its role.

256
00:13:01.075 --> 00:13:03.575
So our first tip is add your persona.

257
00:13:03.575 --> 00:13:05.895
And in this case I'm asking the AI

258
00:13:05.895 --> 00:13:07.215
to be a member of our support team.

259
00:13:07.695 --> 00:13:12.055
'cause my goal is to get the AI to share back

260
00:13:12.075 --> 00:13:14.775
and report to me on our customer feedback.

261
00:13:15.075 --> 00:13:16.215
So this is a really simple,

262
00:13:16.855 --> 00:13:18.735
a simple prompt other than the persona.

263
00:13:18.755 --> 00:13:21.815
And I have my list of customer feedback

264
00:13:21.815 --> 00:13:23.015
here in this bulleted list.

265
00:13:23.395 --> 00:13:26.855
So we're gonna continue building on this prompt as we move

266
00:13:26.855 --> 00:13:27.855
through the webinar tips.

267
00:13:28.835 --> 00:13:31.175
But let's get started by clicking the submit button

268
00:13:31.395 --> 00:13:33.575
and see how the two models do.

269
00:13:34.355 --> 00:13:37.335
So right now what we're doing is the Spark AI playground is

270
00:13:37.335 --> 00:13:39.255
setting an API call to these models.

271
00:13:39.915 --> 00:13:43.335
Now the prompt that I entered will not be shared, um,

272
00:13:43.445 --> 00:13:45.335
back into their, their training data set.

273
00:13:45.755 --> 00:13:47.815
So there's no concern over that.

274
00:13:48.715 --> 00:13:52.255
And you can see Gemini was a little faster than GBT.

275
00:13:52.395 --> 00:13:54.135
So let's take a look at these responses.

276
00:13:55.555 --> 00:13:57.855
So we can see we got a pretty basic report.

277
00:13:58.515 --> 00:14:01.495
Um, it's almost fitting back to me exactly what I had from

278
00:14:01.495 --> 00:14:04.455
that list and I feel like we could do better.

279
00:14:04.865 --> 00:14:06.415
Let's see how Gemini did.

280
00:14:07.775 --> 00:14:11.055
I am getting a report with common issues and some feedback,

281
00:14:11.515 --> 00:14:13.215
but I wanna spice this up a little.

282
00:14:13.375 --> 00:14:14.815
I wanna get a more thorough report.

283
00:14:15.235 --> 00:14:17.655
So let's move on to prompt number two.

284
00:14:24.575 --> 00:14:28.035
I'm just gonna copy and paste it here in my prompt text box.

285
00:14:30.575 --> 00:14:34.635
So my second tip is to tell the model your spec, your

286
00:14:35.195 --> 00:14:36.355
specific expectations.

287
00:14:37.575 --> 00:14:41.075
For example, I may wanna specify that the report be pro uh,

288
00:14:41.105 --> 00:14:42.675
formatted in a table

289
00:14:42.975 --> 00:14:45.755
or have a specified number of paragraphs or word count.

290
00:14:46.495 --> 00:14:48.355
So I like to think about the instructions

291
00:14:48.355 --> 00:14:50.715
that I give my real world teammates when I

292
00:14:50.715 --> 00:14:51.805
hand off a work task.

293
00:14:52.265 --> 00:14:53.685
So I work with a technical writer

294
00:14:54.385 --> 00:14:56.205
and when she starts on a new document,

295
00:14:56.505 --> 00:14:57.685
we have a conversation

296
00:14:58.185 --> 00:15:01.365
and we talk about what the tone should be for that document.

297
00:15:02.905 --> 00:15:06.045
We also talk about how long we want that document to be.

298
00:15:07.065 --> 00:15:10.045
And then for complex documents, we often talk about

299
00:15:10.185 --> 00:15:11.885
how we should organize the information

300
00:15:12.385 --> 00:15:13.765
and section that document.

301
00:15:14.385 --> 00:15:17.445
And I like to do that as well with, with my AI model.

302
00:15:18.225 --> 00:15:21.925
So now that we gave some ex explicit expectations

303
00:15:21.945 --> 00:15:23.885
to the model, let's run this prompt

304
00:15:24.145 --> 00:15:25.685
and see how our two models do.

305
00:15:32.855 --> 00:15:36.825
Okay, again, it looks like Google Gemini won the race.

306
00:15:37.205 --> 00:15:38.225
So let's take a look.

307
00:15:39.085 --> 00:15:41.865
So I can see that I got the table that I asked for

308
00:15:41.865 --> 00:15:42.905
to format in tables.

309
00:15:43.505 --> 00:15:45.265
I am gonna point out that these numbers look

310
00:15:45.495 --> 00:15:46.785
supe suspicious to me.

311
00:15:47.345 --> 00:15:48.865
I would double check them

312
00:15:49.055 --> 00:15:51.505
because this to me looks like a hallucination.

313
00:15:51.845 --> 00:15:53.665
So it looks like Google Gemini is doing

314
00:15:53.665 --> 00:15:54.825
some hallucinating here.

315
00:15:55.805 --> 00:15:57.945
We can see I got the sections that I asked for

316
00:15:59.045 --> 00:16:02.145
and I also asked it to supply a proposed action item.

317
00:16:04.145 --> 00:16:08.255
Let's see how GBT four did also getting my sections.

318
00:16:08.255 --> 00:16:09.895
This is a better formatted report.

319
00:16:10.655 --> 00:16:13.255
I have my table, my frequencies knowing

320
00:16:13.275 --> 00:16:14.415
the data that I sent in.

321
00:16:14.415 --> 00:16:15.535
These look correct to me.

322
00:16:18.875 --> 00:16:20.975
And I also get, I asked for an action item

323
00:16:21.035 --> 00:16:22.175
and it's responding with that.

324
00:16:22.195 --> 00:16:23.895
So you can see this report is already getting a better

325
00:16:24.215 --> 00:16:25.455
structure, a better foundation,

326
00:16:25.555 --> 00:16:27.695
and better returns than my first prompt,

327
00:16:28.475 --> 00:16:30.215
but we can make it even better.

328
00:16:31.195 --> 00:16:32.975
So let's go on to prompt number three.

329
00:16:39.905 --> 00:16:41.835
Okay, this is actually my favorite tip.

330
00:16:45.385 --> 00:16:46.405
So prompt number three.

331
00:16:46.785 --> 00:16:47.805
Uh, my favorite tip is

332
00:16:47.805 --> 00:16:50.485
to organize your prompt into sections.

333
00:16:51.505 --> 00:16:55.445
So my colleagues at ITS know that I often add sections

334
00:16:55.445 --> 00:16:57.165
to my emails and it's

335
00:16:57.165 --> 00:17:00.525
to break up my emails into like logical, logical sections.

336
00:17:00.645 --> 00:17:03.685
I find it's easier for me to organize my idea, my ideas

337
00:17:03.685 --> 00:17:05.605
and make sure I convey everything that I need to convey.

338
00:17:05.985 --> 00:17:08.565
And I like to think it makes it easier on my colleagues

339
00:17:08.585 --> 00:17:10.805
to scan my emails and find what they're looking for.

340
00:17:11.505 --> 00:17:14.605
So I find it works the same way with the, with the AI model.

341
00:17:15.345 --> 00:17:16.605
Now if I'm organized

342
00:17:16.605 --> 00:17:19.485
and I'm giving it all its information, it's more likely to,

343
00:17:19.705 --> 00:17:21.805
to treat that data, um, more equal

344
00:17:21.905 --> 00:17:23.485
and to supply the response I'm looking for.

345
00:17:24.545 --> 00:17:27.765
So I added delimiters and in this case section headers.

346
00:17:27.905 --> 00:17:30.685
So persona, instructions, length structure

347
00:17:30.985 --> 00:17:32.205
to organize my prompt.

348
00:17:32.665 --> 00:17:36.645
But you could also use XML tags or markdown languages

349
00:17:37.305 --> 00:17:39.525
or other, other headers that you see fit.

350
00:17:40.555 --> 00:17:43.175
So I find this helps the generative AI stay on task.

351
00:17:43.585 --> 00:17:44.535
Let's see how they do.

352
00:17:52.305 --> 00:17:55.085
All right, again, Google Gemini won the race.

353
00:17:56.105 --> 00:17:58.925
So this report is getting even more in depth.

354
00:17:59.905 --> 00:18:02.405
So we're getting our introduction section, our summary

355
00:18:02.585 --> 00:18:03.685
of support requests.

356
00:18:04.165 --> 00:18:06.125
I love that it's giving me these frequencies.

357
00:18:06.325 --> 00:18:07.405
I would double check them though,

358
00:18:07.985 --> 00:18:09.405
and descriptions of the issues.

359
00:18:10.545 --> 00:18:13.365
So this is getting enriched, right?

360
00:18:13.965 --> 00:18:17.885
I had also asked in this prompt for it to give me, um, kind

361
00:18:17.885 --> 00:18:19.325
of summarize each section.

362
00:18:19.465 --> 00:18:20.925
And it looks like Gemini did not

363
00:18:20.925 --> 00:18:22.205
respond to that section of the prompt.

364
00:18:22.215 --> 00:18:26.815
Super well, let's see how GPT-4 did.

365
00:18:27.785 --> 00:18:29.095
Again, I have my sections.

366
00:18:30.395 --> 00:18:33.915
I like my frequency counts and it looks like open AI's.

367
00:18:34.235 --> 00:18:37.555
GPT-4 was better at supplying that summary to each section

368
00:18:37.555 --> 00:18:38.795
that I asked it to supply.

369
00:18:39.455 --> 00:18:40.995
So you can see my report is starting

370
00:18:40.995 --> 00:18:43.275
to get more fleshed out and more useful.

371
00:18:43.745 --> 00:18:45.515
This is turning into something I could hand over

372
00:18:45.515 --> 00:18:46.515
to my executive team.

373
00:18:49.325 --> 00:18:50.935
Okay, so prompt number four,

374
00:18:56.455 --> 00:18:58.515
I'm just gonna copy and paste this one in.

375
00:18:59.495 --> 00:19:00.635
And then let's take a look.

376
00:19:03.235 --> 00:19:06.135
So tip number four is to try adding an example

377
00:19:06.315 --> 00:19:07.415
of what you're looking for.

378
00:19:09.115 --> 00:19:13.055
So I tend to use this tip if my earlier prompts are not

379
00:19:13.075 --> 00:19:14.735
as working as well as I would like them to.

380
00:19:14.915 --> 00:19:17.095
I'm just still not getting the format I want.

381
00:19:17.235 --> 00:19:18.615
I'm not getting the tone I want.

382
00:19:18.735 --> 00:19:20.535
I need to give an example to the ai.

383
00:19:21.475 --> 00:19:23.375
So all of our prompts up

384
00:19:23.375 --> 00:19:25.575
to this point have been zero shot prompts,

385
00:19:25.605 --> 00:19:27.295
haven't supplied any examples in them.

386
00:19:28.005 --> 00:19:31.775
This prompt that I just pasted in is a single shot prompt.

387
00:19:32.395 --> 00:19:34.135
So you can see here I have four example

388
00:19:34.755 --> 00:19:36.255
and I actually gave it my example

389
00:19:36.935 --> 00:19:40.215
'cause I wanted to format my action items kind

390
00:19:40.215 --> 00:19:42.415
of in this wording format where it starts

391
00:19:42.415 --> 00:19:43.575
with this action verb.

392
00:19:44.365 --> 00:19:47.695
Okay? So I'm gonna send this, this new prompt

393
00:19:47.725 --> 00:19:49.695
with the example in it over to my models

394
00:19:49.915 --> 00:19:50.935
and see how they perform.

395
00:19:57.625 --> 00:19:59.675
Okay, let's check out on, uh, Gemini.

396
00:20:00.695 --> 00:20:02.075
So again, I get my sections.

397
00:20:02.105 --> 00:20:04.555
This is becoming a really nice fleshed out report

398
00:20:06.405 --> 00:20:08.865
and you can see that my action items are new newly

399
00:20:08.865 --> 00:20:11.105
formatted, using my action verbs to start.

400
00:20:11.415 --> 00:20:13.625
This is closer to what I was hoping for

401
00:20:14.885 --> 00:20:16.945
and it's giving me some interesting action items.

402
00:20:18.395 --> 00:20:20.745
Let's take a look at how GPT-4 did.

403
00:20:25.215 --> 00:20:26.635
So again, I get my sections.

404
00:20:28.275 --> 00:20:29.615
I'm scrolling down my action items.

405
00:20:30.365 --> 00:20:32.775
This was my exact example

406
00:20:33.755 --> 00:20:35.975
and I'm happy that GPT-4

407
00:20:36.575 --> 00:20:38.735
returned my exact example for this reason.

408
00:20:39.355 --> 00:20:41.535
One of the reasons that I tend

409
00:20:41.535 --> 00:20:44.895
to not supply an example in my initial prompts is

410
00:20:44.895 --> 00:20:47.095
because sometimes they will bleed over into

411
00:20:47.265 --> 00:20:48.485
the model's response.

412
00:20:49.305 --> 00:20:51.325
And what I mean by that is it duplicates.

413
00:20:51.385 --> 00:20:54.285
My example in the response, I didn't want GPT-4

414
00:20:54.285 --> 00:20:55.445
to duplicate my example.

415
00:20:55.605 --> 00:20:57.605
I wanted it to gimme some new action items,

416
00:20:57.775 --> 00:21:00.685
insightful action items, um, instead

417
00:21:00.685 --> 00:21:03.285
of just repeating myself to me for verbatim.

418
00:21:04.185 --> 00:21:06.565
So that's a little disappointing from GPT-4,

419
00:21:06.625 --> 00:21:09.085
but you get to see how the two models reacted to

420
00:21:09.085 --> 00:21:10.125
that example differently.

421
00:21:10.865 --> 00:21:13.605
So I tend to use examples for just one-off prompts

422
00:21:13.605 --> 00:21:15.845
that I don't plan to reuse if I use them.

423
00:21:16.185 --> 00:21:18.885
Um, I am weary of putting them in prompts

424
00:21:18.885 --> 00:21:21.365
that get reused repeatedly, especially ones

425
00:21:21.365 --> 00:21:22.765
that get repeated by programs.

426
00:21:24.045 --> 00:21:28.645
I, okay. Okay, I'm gonna move on

427
00:21:28.645 --> 00:21:29.885
to our fifth prompt.

428
00:21:33.865 --> 00:21:36.675
Just gonna copy and paste this in so that we can all see it.

429
00:21:50.105 --> 00:21:52.795
Okay, tip number five, I want you

430
00:21:52.795 --> 00:21:55.155
to think about back when you were onboarding a

431
00:21:55.155 --> 00:21:56.355
mem a new member to your team.

432
00:21:57.985 --> 00:21:59.395
When you onboard a new member

433
00:22:00.095 --> 00:22:01.915
and they're given a complex task,

434
00:22:01.915 --> 00:22:03.795
they're often not sure how to get started.

435
00:22:03.985 --> 00:22:05.115
It's very daunting.

436
00:22:06.055 --> 00:22:08.795
And what I do for new team members on my team is I start

437
00:22:09.075 --> 00:22:12.475
breaking up those, those complex tasks into smaller steps.

438
00:22:13.295 --> 00:22:17.315
So you can do that as well for your prompt here to the AI

439
00:22:17.975 --> 00:22:20.395
in this, in this case, I'm breaking telling it

440
00:22:20.395 --> 00:22:22.235
to follow these steps to write my report.

441
00:22:22.435 --> 00:22:25.115
I wanted to categorize the requests and the feedback.

442
00:22:25.835 --> 00:22:27.435
I wanted to count the frequencies.

443
00:22:28.955 --> 00:22:31.275
I like number three, I wanted to highlight the most

444
00:22:31.835 --> 00:22:33.115
frequent support requests

445
00:22:33.255 --> 00:22:36.275
and share with me any patterns that it's observing.

446
00:22:37.065 --> 00:22:39.395
Okay? I wanna assess the potential impact

447
00:22:39.455 --> 00:22:41.675
of each issue on the user experience.

448
00:22:42.955 --> 00:22:45.375
And I wanna separate the feedback into positive

449
00:22:45.475 --> 00:22:49.645
and negative pieces of feedback. Basically,

450
00:22:49.705 --> 00:22:52.645
Andrea, on this one, essentially the step-by-step is like

451
00:22:53.255 --> 00:22:56.885
debriefing a team member on all the context they need

452
00:22:56.885 --> 00:22:59.245
to give you exactly what you want, right? Yeah.

453
00:22:59.315 --> 00:23:00.485
It's getting really detailed.

454
00:23:00.635 --> 00:23:02.765
Like these are, these are the things that I'm hoping

455
00:23:02.785 --> 00:23:05.045
to see in this output so that I'm,

456
00:23:05.405 --> 00:23:06.565
'cause it's what I need in my report

457
00:23:06.785 --> 00:23:08.925
and it's what's gonna satisfy the original request.

458
00:23:09.185 --> 00:23:11.325
So it's really breaking it down into steps

459
00:23:11.345 --> 00:23:13.645
so it's more achievable, um,

460
00:23:13.745 --> 00:23:15.685
and leaves less to to chance here.

461
00:23:16.105 --> 00:23:18.725
So another term for this kind of prompting, Amanda,

462
00:23:19.105 --> 00:23:20.605
is chain of thought prompting.

463
00:23:20.745 --> 00:23:22.805
So if you ever hear that term when you're looking up,

464
00:23:22.805 --> 00:23:26.485
prompting, prompting tips, um, asking, giving it step

465
00:23:26.485 --> 00:23:27.485
by step instructions.

466
00:23:27.485 --> 00:23:29.485
It's called chain of thought prompting.

467
00:23:30.985 --> 00:23:34.045
So we also have a question. So for this one mm-Hmm.

468
00:23:34.125 --> 00:23:36.485
Um, someone asked, would you please comment

469
00:23:36.665 --> 00:23:39.725
how the configuration options to the right may be used?

470
00:23:39.945 --> 00:23:43.045
So if you wanted to walk through some of that as well.

471
00:23:43.135 --> 00:23:45.885
Great. Absolutely. Um, I, I can share some of that.

472
00:23:46.265 --> 00:23:49.725
Um, can I, Amanda, I'm gonna finish first with showing how,

473
00:23:49.905 --> 00:23:52.445
how the prompt the AI response to this prompt,

474
00:23:52.545 --> 00:23:53.725
and then we can talk about those

475
00:23:53.965 --> 00:23:55.285
configuration options on the right.

476
00:23:55.545 --> 00:23:57.965
Sounds good. Okay. Okay.

477
00:23:57.965 --> 00:24:00.165
So I'm gonna send this over to the, to the ai.

478
00:24:00.185 --> 00:24:02.365
And one of the reasons I really like chain

479
00:24:02.385 --> 00:24:04.165
of thought prompting is

480
00:24:04.165 --> 00:24:06.365
because ai, like I'm sending them the prompts,

481
00:24:06.525 --> 00:24:08.645
I don't know what's happening on happening over there,

482
00:24:08.665 --> 00:24:10.205
and then it's sending me this response.

483
00:24:10.225 --> 00:24:11.645
It feels almost like a black box,

484
00:24:12.065 --> 00:24:14.205
but when I'm able to tell it how I want it to think

485
00:24:14.205 --> 00:24:17.125
through the problem, I have a bit of an idea of

486
00:24:17.225 --> 00:24:18.525
how it's gonna work through the problem

487
00:24:18.545 --> 00:24:20.085
and return that response to me.

488
00:24:20.705 --> 00:24:25.565
Um, all right, let's start with our GPT-4 response.

489
00:24:26.595 --> 00:24:28.605
Okay. We can see I got my frequencies

490
00:24:29.545 --> 00:24:32.365
and you can see I have my an area now explaining the most

491
00:24:32.845 --> 00:24:35.485
frequent support requests and the patterns I was asking for,

492
00:24:35.705 --> 00:24:39.365
and also the impact of those issues on the user experience.

493
00:24:40.025 --> 00:24:41.725
So I'm getting a more meaningful

494
00:24:41.725 --> 00:24:43.485
and enriched report that I can really share

495
00:24:43.485 --> 00:24:47.445
with my executives and move things forward.

496
00:24:48.825 --> 00:24:50.925
So same with here with Google Gemini,

497
00:24:51.145 --> 00:24:53.925
you can see it really responded back with the summaries.

498
00:24:54.595 --> 00:24:56.125
It's highlighting the most frequent issues

499
00:24:56.125 --> 00:24:57.565
and patterns as I asked it to,

500
00:24:58.705 --> 00:25:00.845
and it's responding with that user experience

501
00:25:00.845 --> 00:25:03.605
that I wasn't getting with my earlier prompts.

502
00:25:04.595 --> 00:25:06.325
Okay, so let's shift over to some

503
00:25:06.325 --> 00:25:08.645
of these settings here on here on the right.

504
00:25:09.105 --> 00:25:12.285
So over in the playground, we expose these settings for you.

505
00:25:12.305 --> 00:25:13.405
So you have a lot of control

506
00:25:13.405 --> 00:25:15.165
and you can play with them to see how it changes

507
00:25:15.465 --> 00:25:19.405
how the model responds to your, to your prompt.

508
00:25:19.905 --> 00:25:21.605
And you're gonna see the settings are not

509
00:25:21.605 --> 00:25:22.845
identical for all the models.

510
00:25:22.945 --> 00:25:24.605
So it really depends on what the provider

511
00:25:24.625 --> 00:25:25.685
and the model support.

512
00:25:26.865 --> 00:25:31.485
So temperature, temperature directly controls randomness.

513
00:25:32.225 --> 00:25:35.205
So you can think of randomness as akin to creativity.

514
00:25:35.825 --> 00:25:38.725
How creative is the AI model gonna be in

515
00:25:38.885 --> 00:25:40.325
choosing the next word?

516
00:25:40.795 --> 00:25:43.525
Okay, so a higher temperature, if I move this higher,

517
00:25:44.055 --> 00:25:46.005
we're gonna get more creative.

518
00:25:46.595 --> 00:25:50.085
Okay? And if I keep it really kinda low, it's going

519
00:25:50.085 --> 00:25:53.045
to be less creative and it's gonna give more predictable

520
00:25:53.425 --> 00:25:54.845
and consistent output.

521
00:25:55.795 --> 00:25:58.765
Okay? So now top P is a little different.

522
00:25:59.065 --> 00:26:01.205
Um, generally I don't recommend using both

523
00:26:01.205 --> 00:26:02.605
temperature and top P together.

524
00:26:03.225 --> 00:26:07.765
Um, top P tells the model kind of how to select a pool

525
00:26:07.945 --> 00:26:11.565
of next words and then select the next word from that pool.

526
00:26:12.145 --> 00:26:13.765
So that's how it's handling kind

527
00:26:13.765 --> 00:26:15.245
of creativity and randomness.

528
00:26:16.965 --> 00:26:20.545
Now, this presence penalty, this is to encourage the model

529
00:26:20.605 --> 00:26:23.225
to include a more diverse range of tokens.

530
00:26:23.245 --> 00:26:26.705
Tokens are like three-fourths a word, and a higher value.

531
00:26:26.895 --> 00:26:28.905
It's gonna result in the model being more likely

532
00:26:28.905 --> 00:26:31.665
to generate, um, tokens.

533
00:26:31.685 --> 00:26:33.305
So words that haven't been used

534
00:26:33.705 --> 00:26:35.025
previously within the response.

535
00:26:36.585 --> 00:26:38.805
And then the frequency penalty.

536
00:26:40.065 --> 00:26:42.005
Um, the frequency penalty that's used

537
00:26:42.005 --> 00:26:44.125
to discourage the model from repeating the same words

538
00:26:44.125 --> 00:26:47.125
and phrases too often within the generated response.

539
00:26:47.665 --> 00:26:51.005
So a higher value is gonna result in the model being more

540
00:26:51.085 --> 00:26:54.365
conservative in its use of repeating the tokens.

541
00:26:54.745 --> 00:26:57.805
So there's a lot of overlap in how, how these things work.

542
00:26:57.865 --> 00:27:01.365
And really you can play with them to keep refining, um,

543
00:27:01.665 --> 00:27:03.445
how the quality of that response.

544
00:27:04.745 --> 00:27:05.765
Now my favorite,

545
00:27:06.585 --> 00:27:09.405
my favorite setting over here is number of responses.

546
00:27:09.865 --> 00:27:13.125
So I can change the number of responses and send a prompt.

547
00:27:13.265 --> 00:27:18.165
And if I have it set to two, I'm asking GBT four to send me,

548
00:27:19.385 --> 00:27:21.845
uh, to send two responses.

549
00:27:22.465 --> 00:27:24.885
And I can see how the two different responses differ.

550
00:27:25.665 --> 00:27:30.225
So I'm gonna copy

551
00:27:30.225 --> 00:27:31.465
and paste my last prompt.

552
00:27:33.915 --> 00:27:36.535
And quickly just send this over and ask for two responses

553
00:27:36.535 --> 00:27:39.055
and you'll see how I get two responses from g pt

554
00:27:39.055 --> 00:27:40.575
four to the same prompt.

555
00:27:51.085 --> 00:27:53.415
Okay. All right. So here's my first one.

556
00:27:53.615 --> 00:27:55.735
I think it's just taking a moment to return the second one,

557
00:27:55.735 --> 00:27:57.015
but it'll, when it's done, it should

558
00:27:57.015 --> 00:27:58.135
appear right here beneath.

559
00:27:58.835 --> 00:28:03.575
Um, Amanda, do we have time for some bonus, bonus tips

560
00:28:03.575 --> 00:28:05.055
after I give a quick run through

561
00:28:05.055 --> 00:28:06.535
of the five tips we went through already?

562
00:28:07.395 --> 00:28:10.815
We have about 10 minutes left for your section. Mm-Hmm.

563
00:28:11.105 --> 00:28:13.655
Great. So just to recap our five tips today,

564
00:28:13.825 --> 00:28:18.615
we're adding a persona, getting specific, adding delimiters

565
00:28:18.615 --> 00:28:21.655
and sections to your prompt, providing an example.

566
00:28:22.475 --> 00:28:25.695
And then our last one was listing out the steps you would

567
00:28:25.695 --> 00:28:27.295
like the generative AI model to take.

568
00:28:28.155 --> 00:28:30.175
So I'm gonna stop sharing my screen

569
00:28:37.045 --> 00:28:38.825
Did have a question about models,

570
00:28:38.875 --> 00:28:40.305
which I know you're gonna get to.

571
00:28:40.485 --> 00:28:42.385
And, uh, but they ask, curious

572
00:28:42.445 --> 00:28:44.745
to know if you have a favorite or most used model.

573
00:28:45.535 --> 00:28:50.185
Okay, I do. Um, GPT-4 Omni is my favorite model right now.

574
00:28:50.565 --> 00:28:54.155
Um, you know, Amanda, I'm having a hard time figuring out

575
00:28:54.215 --> 00:28:55.435
how to stop sharing.

576
00:28:56.695 --> 00:28:58.355
Uh, I think it's at the top of the screen.

577
00:28:58.505 --> 00:29:00.315
It's should I be? Thank you,

578
00:29:00.805 --> 00:29:01.805
Thank you. I per,

579
00:29:01.805 --> 00:29:02.355

580
00:29:02.645 --> 00:29:04.795
We're not a team organization, I mean,

581
00:29:04.825 --> 00:29:06.995
assumes the organization's. No worries. No,

582
00:29:07.335 --> 00:29:08.335
No. Okay.

583
00:29:08.335 --> 00:29:10.315
So I have some bonus tips today. Okay.

584
00:29:10.775 --> 00:29:14.755
Um, one is you may not wanna have to outline all the steps

585
00:29:14.775 --> 00:29:15.995
for the generative AI model,

586
00:29:16.015 --> 00:29:17.715
but you might really want it to think it through.

587
00:29:17.935 --> 00:29:20.155
So you can just simply add wording to your prompt.

588
00:29:20.155 --> 00:29:22.995
That is something like think through this step by step

589
00:29:23.095 --> 00:29:24.475
and the model's gonna slow down

590
00:29:24.735 --> 00:29:27.035
and it's gonna share with you its thought process,

591
00:29:27.295 --> 00:29:30.115
and it's gonna work through it in a, in a more kind

592
00:29:30.115 --> 00:29:34.195
of defined and methodol methodological kind of way.

593
00:29:35.415 --> 00:29:38.125
Bonus tip number two is when I write

594
00:29:38.125 --> 00:29:39.525
emails, I don't bury the lead.

595
00:29:39.605 --> 00:29:41.045
I put the most important piece first.

596
00:29:41.395 --> 00:29:43.125
Sometimes I repeat it at the end.

597
00:29:43.505 --> 00:29:44.845
You can do the same with your model.

598
00:29:44.845 --> 00:29:47.605
If you wanna make sure it does not miss this piece, it,

599
00:29:48.235 --> 00:29:49.565
it's going to concentrate there.

600
00:29:51.415 --> 00:29:53.315
Tip number three, and it's my last bonus tip.

601
00:29:54.265 --> 00:29:55.795
It's continue the conversation.

602
00:29:56.415 --> 00:29:58.275
So you do not have to send a mega prompt,

603
00:29:58.435 --> 00:30:00.635
a really large prompt with everything in it the way

604
00:30:00.635 --> 00:30:02.715
that we were doing today, just for speed.

605
00:30:03.335 --> 00:30:06.515
Um, you can break it up into smaller prompts

606
00:30:06.895 --> 00:30:10.435
and kind of continue informing the model to make tweaks

607
00:30:10.495 --> 00:30:13.155
to the same output so you can have the model kind

608
00:30:13.155 --> 00:30:14.395
of just keep amending and adding

609
00:30:14.415 --> 00:30:15.595
to the output that it's creating.

610
00:30:16.655 --> 00:30:19.195
Amanda, do we have time to go into some model comparison

611
00:30:19.255 --> 00:30:21.235
to answer some of those questions that were coming through?

612
00:30:21.895 --> 00:30:24.235
We do. And we have a question. I really like this one.

613
00:30:24.375 --> 00:30:27.915
So, uh, John asked, how specific do you find it useful

614
00:30:27.975 --> 00:30:30.395
to get in terms of the context persona

615
00:30:30.395 --> 00:30:31.515
at the beginning of the prompt?

616
00:30:33.055 --> 00:30:35.015
I find it really useful. Yeah.

617
00:30:35.035 --> 00:30:38.555
Um, so for example,

618
00:30:40.565 --> 00:30:45.105
if I fed, if I fed that generative AI a piece of text

619
00:30:45.165 --> 00:30:49.625
and I said edit, edit this text, they may rewrite it,

620
00:30:49.775 --> 00:30:51.625
they may take awkward sentences and make them better.

621
00:30:51.625 --> 00:30:53.705
But you know what? I had specific wording in there.

622
00:30:53.985 --> 00:30:55.985
I had things exactly the way I wanted it.

623
00:30:56.265 --> 00:30:58.345
I really just wanted it to find my grammar issues.

624
00:30:58.925 --> 00:31:00.745
Um, and I find by giving its persona

625
00:31:00.745 --> 00:31:02.465
and telling its role, it's going to,

626
00:31:02.575 --> 00:31:05.825
it's gonna do a better job at meeting my expectations. Um,

627
00:31:06.825 --> 00:31:07.825
I definitely agree.

628
00:31:08.365 --> 00:31:10.025
Um, yeah, we do have time, uh,

629
00:31:10.025 --> 00:31:11.825
we've got about 10 minutes if you wanna go

630
00:31:11.825 --> 00:31:12.985
through model comparisons.

631
00:31:13.615 --> 00:31:14.825
Yeah. Okay.

632
00:31:15.165 --> 00:31:18.145
So we just saw in real time how different models responded

633
00:31:18.145 --> 00:31:21.545
to the same prompts in that, in the Spark AI playground.

634
00:31:22.245 --> 00:31:24.575
Um, I'd like to talk about some of those,

635
00:31:24.575 --> 00:31:27.095
those main differences and models fall into kind

636
00:31:27.095 --> 00:31:28.415
of three major categories.

637
00:31:28.955 --> 00:31:32.415
So the first is decoder models, and that's like GPT,

638
00:31:32.955 --> 00:31:35.095
and they're really good at generating content.

639
00:31:35.635 --> 00:31:39.015
So why is GPT-4 omni my favorite model right now?

640
00:31:40.035 --> 00:31:43.175
I'm finding it's just really good at text completion tasks,

641
00:31:43.175 --> 00:31:45.815
especially generating test questions that I need

642
00:31:45.815 --> 00:31:47.575
for like demos that I'm doing for clients.

643
00:31:47.955 --> 00:31:50.695
So it's really my go-to for generating content.

644
00:31:52.275 --> 00:31:55.935
So also within the decoder kind of category of models,

645
00:31:56.555 --> 00:31:59.615
you have models like AI 21 Labs Jurassic,

646
00:31:59.615 --> 00:32:01.255
and I'm hearing really good things about

647
00:32:01.255 --> 00:32:02.735
that model for translation.

648
00:32:03.995 --> 00:32:07.615
So you also have, um, anthros, Claude

649
00:32:07.795 --> 00:32:09.655
and I did have a, you know,

650
00:32:10.175 --> 00:32:12.935
feedback from a colleague on Claude sharing

651
00:32:12.935 --> 00:32:15.255
that this model was particularly good at creative writing.

652
00:32:15.395 --> 00:32:18.215
So they were seeing great uses for it in the K to 12 space

653
00:32:18.215 --> 00:32:19.535
where you're writing kind of creative

654
00:32:19.535 --> 00:32:20.855
writing for reading comprehension.

655
00:32:22.085 --> 00:32:23.415
Then you have another category

656
00:32:23.415 --> 00:32:25.175
of models, um, encoder models.

657
00:32:25.795 --> 00:32:27.575
And these models like Bert,

658
00:32:28.205 --> 00:32:30.415
they're really good at understanding relationships

659
00:32:30.875 --> 00:32:32.935
and they're, they're models you may wanna look into if

660
00:32:32.935 --> 00:32:35.735
you're looking to analyze data and categorize data

661
00:32:35.735 --> 00:32:38.495
and classify data so they have strengths in that area.

662
00:32:39.865 --> 00:32:43.405
The third main category is your encoder decoder models.

663
00:32:43.435 --> 00:32:46.645
They're both, so they have strengths in both kind of

664
00:32:46.645 --> 00:32:48.925
that noticing relationships and generating content.

665
00:32:49.505 --> 00:32:51.405
And these are models like Google Gemini,

666
00:32:51.405 --> 00:32:52.525
which you saw working today.

667
00:32:53.895 --> 00:32:56.915
Um, now AI providers, they are eager

668
00:32:56.915 --> 00:32:58.315
to tell you about their models.

669
00:32:58.655 --> 00:33:00.275
So if you just Google the provider

670
00:33:00.275 --> 00:33:02.435
and you go to their main site, they're gonna tell you

671
00:33:02.435 --> 00:33:04.195
what they think, each model that they,

672
00:33:04.385 --> 00:33:06.835
they have on offer is particularly good at.

673
00:33:07.655 --> 00:33:10.235
And then after you read that, come to the AI playground

674
00:33:10.575 --> 00:33:11.795
and then select those models

675
00:33:11.935 --> 00:33:16.035
and see how they, how they compare when you send them your

676
00:33:16.275 --> 00:33:17.475
objectives and your tasks.

677
00:33:17.535 --> 00:33:20.595
So which model is the best for your purposes?

678
00:33:21.575 --> 00:33:23.515
So it's one of the great benefits

679
00:33:23.515 --> 00:33:26.035
of having the AI playground available both internally

680
00:33:26.215 --> 00:33:30.515
and now externally to our clients after performance.

681
00:33:30.515 --> 00:33:31.755
There's a couple other things that,

682
00:33:31.755 --> 00:33:32.795
that you may wanna consider.

683
00:33:33.255 --> 00:33:34.915
Um, one would be cost.

684
00:33:35.295 --> 00:33:38.555
So the, the AI providers do charge a cost

685
00:33:38.815 --> 00:33:40.195
for using their models,

686
00:33:40.615 --> 00:33:43.795
and they charge a cost for the size of the input.

687
00:33:43.815 --> 00:33:44.915
So the size of your prompt,

688
00:33:45.375 --> 00:33:47.875
and then also the size of the response that's coming back.

689
00:33:48.655 --> 00:33:52.915
Now, what we found, um, working with models for a IG is

690
00:33:52.915 --> 00:33:54.475
that the cost range for the models

691
00:33:54.475 --> 00:33:56.115
that we support wasn't that huge.

692
00:33:56.615 --> 00:33:59.035
Um, one of my colleagues, Chris Glackin, found

693
00:33:59.105 --> 00:34:01.995
that we could generate 2000 multiple choice questions

694
00:34:01.995 --> 00:34:03.395
for less than $32.

695
00:34:04.065 --> 00:34:05.915
Okay? So I, I personally find cost

696
00:34:05.915 --> 00:34:07.315
as a lower, lower concern.

697
00:34:08.795 --> 00:34:11.135
Now, these different providers provide different support

698
00:34:11.175 --> 00:34:13.815
toolings, so it's something to go look at on their website.

699
00:34:14.475 --> 00:34:17.695
Um, you may be interested in say fine tuning a model,

700
00:34:17.915 --> 00:34:21.095
and some of these providers have tools to help you do that.

701
00:34:21.355 --> 00:34:23.095
So what is fine tuning a model?

702
00:34:23.685 --> 00:34:26.095
Fine tuning a model is when you have a data set,

703
00:34:26.355 --> 00:34:29.935
an extra data set, that's your domain, that's your data,

704
00:34:30.755 --> 00:34:32.495
and you want to feed it to one

705
00:34:32.495 --> 00:34:37.135
of these foundation models like GBT four or Gemini.

706
00:34:37.675 --> 00:34:39.815
So you wanna do further training on those models so

707
00:34:39.815 --> 00:34:43.735
that it can specialize in your tasks and your terminology.

708
00:34:44.765 --> 00:34:48.095
Okay. So there's different costs for using those tools

709
00:34:48.355 --> 00:34:49.655
and fine tuning a model

710
00:34:49.675 --> 00:34:51.655
and those fine tuned models, they then become

711
00:34:51.655 --> 00:34:52.895
available just for your use.

712
00:34:53.355 --> 00:34:56.135
So Amanda, didn't our colleagues do a, a webinar

713
00:34:56.855 --> 00:35:01.565
recently on fine tuning and rag and those ideas? Yeah, two,

714
00:35:01.665 --> 00:35:05.085
Two weeks ago, um, they led one, one rag and fine tuning

715
00:35:05.185 --> 00:35:06.765
and it was super in depth.

716
00:35:07.265 --> 00:35:10.125
Um, that is on the website if anyone wants to revisit that.

717
00:35:10.145 --> 00:35:12.685
And they show the demo of how to actually do it,

718
00:35:13.225 --> 00:35:15.005
um, which is really cool.

719
00:35:15.305 --> 00:35:17.165
You do have a question in the chat if

720
00:35:17.165 --> 00:35:18.525
you want to answer this.

721
00:35:18.615 --> 00:35:20.485
We've got four repetitive tasks.

722
00:35:21.025 --> 00:35:23.565
Do you recom do you recommend on how

723
00:35:23.585 --> 00:35:26.965
to retain the prompt instructions for multiple requests?

724
00:35:27.705 --> 00:35:31.365
How to retain it? Yeah. Okay. Oh, that's a good question.

725
00:35:32.505 --> 00:35:33.845
That's a really good question.

726
00:35:34.545 --> 00:35:37.325
Um, so it depends what that repetitive task is

727
00:35:37.325 --> 00:35:38.445
and who's calling it.

728
00:35:38.865 --> 00:35:42.405
Um, so I know I work, I work a, a set of product managers

729
00:35:42.665 --> 00:35:46.605
and technical writer that keep a notebook of their tasks

730
00:35:47.225 --> 00:35:50.525
and they even write down how the response was.

731
00:35:50.585 --> 00:35:51.685
So they have their preferred tasks

732
00:35:51.685 --> 00:35:53.165
and the ones that are, uh, prompts

733
00:35:53.165 --> 00:35:55.605
and what's working for them so they can go back to them.

734
00:35:55.785 --> 00:35:57.285
So that's at an individual level.

735
00:35:58.025 --> 00:36:00.125
Um, and then organizations, I know a lot

736
00:36:00.125 --> 00:36:02.485
of them are coming up with their own handbooks

737
00:36:02.585 --> 00:36:05.205
and support materials, so they may have some input in

738
00:36:05.205 --> 00:36:08.445
how they want you to store off your, your prompts

739
00:36:08.745 --> 00:36:11.445
and share them with the remainder of your organization. Um,

740
00:36:11.635 --> 00:36:15.445
Yeah, from, um, in marketing we have a, a prompt library,

741
00:36:15.655 --> 00:36:16.805
which sounds really fancy,

742
00:36:16.945 --> 00:36:19.045
but it's just an Excel sheet sheet that says,

743
00:36:19.105 --> 00:36:22.525
here's the prompt we used, here's the use case for it so

744
00:36:22.525 --> 00:36:24.805
that we can come back if we like it, you know,

745
00:36:24.805 --> 00:36:26.485
if it worked well to use it again.

746
00:36:26.865 --> 00:36:29.205
Um, so we're not always starting from scratch, especially

747
00:36:29.205 --> 00:36:30.445
for those longer prompts.

748
00:36:34.025 --> 00:36:36.045
That's a great idea. Amanda, there

749
00:36:36.045 --> 00:36:37.605
Was another question that I'm not totally

750
00:36:37.755 --> 00:36:39.245
sure, um, how to answer.

751
00:36:39.425 --> 00:36:41.725
So I will pass it to you to see what you think.

752
00:36:42.225 --> 00:36:45.005
Um, are there plans to allow users

753
00:36:45.265 --> 00:36:48.565
to use report data from program workshop, um,

754
00:36:48.665 --> 00:36:50.125
in the AI playground?

755
00:36:50.375 --> 00:36:51.485
Which is a great question.

756
00:36:51.595 --> 00:36:54.285
Okay, that's, that's an incredible question. Yeah.

757
00:36:54.585 --> 00:36:56.205
Um, I am gonna have to share

758
00:36:56.205 --> 00:36:58.125
that I am an item workshop product manager,

759
00:36:58.305 --> 00:37:01.325
so I don't work on program workshop, so I can make sure

760
00:37:01.325 --> 00:37:02.845
that question gets over to,

761
00:37:03.345 --> 00:37:05.085
to the product managers in my department

762
00:37:05.085 --> 00:37:06.245
that work on that application.

763
00:37:06.645 --> 00:37:10.065
'cause to be frank, I'm not sure on the specifics

764
00:37:10.065 --> 00:37:12.305
of their roadmap and what they have planning.

765
00:37:15.515 --> 00:37:16.575
That's great. Those are the

766
00:37:16.575 --> 00:37:17.855
two questions we have right now.

767
00:37:18.515 --> 00:37:20.655
Um, we do have a few more minutes if there

768
00:37:20.655 --> 00:37:21.935
was anything else you wanted to cover.

769
00:37:22.395 --> 00:37:24.495
Um, otherwise we can open it up to q and a.

770
00:37:24.515 --> 00:37:25.695
If you have a question,

771
00:37:25.695 --> 00:37:27.695
you can either reuse the raise your hand feature,

772
00:37:27.705 --> 00:37:31.135
which should be available to you, um, or put it in the chat.

773
00:37:32.105 --> 00:37:34.775
Thank you, Amanda. This has been so much fun. This

774
00:37:34.775 --> 00:37:35.775
Has been really great.

775
00:37:36.235 --> 00:37:38.935
Um, was there anything else that you maybe wanted

776
00:37:38.935 --> 00:37:40.415
to cover today or,

777
00:37:40.475 --> 00:37:41.975
it seems like we went through a whole lot.

778
00:37:43.045 --> 00:37:45.095
Yeah, I'm, I think I shared everything

779
00:37:45.095 --> 00:37:46.175
that I had on my mind today,

780
00:37:46.175 --> 00:37:47.535
so I'd love to know if there's any questions,

781
00:37:49.845 --> 00:37:50.845
Hand raises.

782
00:37:52.085 --> 00:37:53.415
Okay. No.

783
00:37:53.925 --> 00:37:56.495
Well then with that I will just do a quick closing

784
00:37:56.635 --> 00:37:59.295
and say thank you to everyone for being here with us.

785
00:37:59.755 --> 00:38:02.375
Uh, we do have part four of our sessions,

786
00:38:02.375 --> 00:38:04.375
our live sessions going live next week.

787
00:38:04.725 --> 00:38:06.255
It's August 6th at 12.

788
00:38:06.725 --> 00:38:08.775
That session is on data analysis

789
00:38:08.955 --> 00:38:10.735
and optimizing survey results.

790
00:38:10.955 --> 00:38:13.015
So if you have candidates that are taking a survey,

791
00:38:13.155 --> 00:38:14.735
you really don't wanna miss this one.

792
00:38:15.195 --> 00:38:16.655
Um, it is always recorded,

793
00:38:16.755 --> 00:38:18.295
but we hope that you join us live,

794
00:38:18.295 --> 00:38:19.895
especially if you have any questions.

795
00:38:20.435 --> 00:38:22.815
Um, that's with Jeffrey Za, he's our director

796
00:38:22.815 --> 00:38:26.215
of product management and create both our product manager,

797
00:38:27.035 --> 00:38:31.655
um, you would visit testis.com/webinars to sign up,

798
00:38:31.955 --> 00:38:34.255
um, and see all the previous recordings as well.

799
00:38:35.105 --> 00:38:38.805
So with that, um, oh, we had a good thank you.

800
00:38:38.985 --> 00:38:40.805
You're welcome. Mark, thanks for joining us.

801
00:38:41.425 --> 00:38:44.085
Um, and again, if you have questions, just reach out.

802
00:38:44.465 --> 00:38:46.285
Uh, thank you for being here with us today,

803
00:38:46.345 --> 00:38:47.845
and we will see you at the next session.

804
00:38:47.935 --> 00:38:51.485
Thank you, Andrea. Thank you. Bye everyone.