prompt_eng_07_expanding_v3.srt

1
00:00:05,033 --> 00:00:07,966
Expanding as a torso, taking a shorter
piece of text

2
00:00:08,266 --> 00:00:10,933
such as a set of instructions
or a list of topics,

3
00:00:11,166 --> 00:00:14,200
and having the large language model
generate a longer

4
00:00:14,200 --> 00:00:17,966
piece of text such as an email
or an essay about some topic.

5
00:00:18,433 --> 00:00:21,900
There are some great uses of this, such
as if you use the large language model

6
00:00:22,100 --> 00:00:24,066
as a brainstorming partner.

7
00:00:24,066 --> 00:00:27,933
But I just also want acknowledge
that there's some problematic use cases

8
00:00:27,933 --> 00:00:32,200
of this, such as if someone with a user
to generate a large amount of spam.

9
00:00:32,666 --> 00:00:35,900
So when you use these capabilities
of a large language model,

10
00:00:36,100 --> 00:00:40,866
please use it only in a responsible way
and in a way that helps people.

11
00:00:41,633 --> 00:00:45,933
In this video, we'll go through an example
of how you can use a language model

12
00:00:45,933 --> 00:00:50,100
to generate a personalized email
based on some information.

13
00:00:50,100 --> 00:00:53,166
The email is kind of self
proclaimed to be from an AI bot,

14
00:00:53,166 --> 00:00:55,800
which as I mentioned, is very important.

15
00:00:56,266 --> 00:00:57,800
What else are you going to use?

16
00:00:57,800 --> 00:01:01,066
Another one of the models
input parameters called temperature,

17
00:01:01,200 --> 00:01:05,666
and this kind of allows you to vary
the kind of degree of exploration

18
00:01:05,666 --> 00:01:08,533
and variety
in the kind of models responses.

19
00:01:08,800 --> 00:01:11,333
So let's get into it.

20
00:01:11,333 --> 00:01:15,166
So before we get started,
we're going to kind of do the usual setup.

21
00:01:15,166 --> 00:01:19,166
So set up the open Python package
and then also define our helper

22
00:01:19,166 --> 00:01:20,433
function, get completion.

23
00:01:22,266 --> 00:01:22,733
And now we're

24
00:01:22,733 --> 00:01:25,500
going to write a custom
email response to a customer review.

25
00:01:25,900 --> 00:01:29,366
And so given a customer view
and the sentiment,

26
00:01:29,366 --> 00:01:31,633
we're going to generate a custom response.

27
00:01:32,366 --> 00:01:35,100
Now we're going to use the language model

28
00:01:35,100 --> 00:01:38,633
to generate a custom email to a customer

29
00:01:39,066 --> 00:01:42,633
based on a customer review
and the sentiment of the review.

30
00:01:43,033 --> 00:01:46,233
So we've already extracted the sentiment
using

31
00:01:46,666 --> 00:01:50,066
the kind of prompts that we saw
in the empowering video.

32
00:01:51,066 --> 00:01:55,900
And then this is the customer
review for a blunder.

33
00:01:55,900 --> 00:01:59,733
And now we're going to customize
the reply based on the sentiment.

34
00:02:01,266 --> 00:02:03,166
And so here the instruction is

35
00:02:03,166 --> 00:02:06,066
your you are our customer
service assistant.

36
00:02:06,266 --> 00:02:08,966
Your task is to send an email reply
to a valued customer.

37
00:02:09,266 --> 00:02:12,900
Given the customer
email is limited by three back to generate

38
00:02:12,900 --> 00:02:14,900
a reply
to thank the customer for that review.

39
00:02:14,900 --> 00:02:18,133
If the sentiment is positive or neutral,
thank them for that review.

40
00:02:18,466 --> 00:02:21,800
If the sentiment is negative, apologize
and suggest that they can reach out

41
00:02:21,800 --> 00:02:23,333
to customer service.

42
00:02:23,333 --> 00:02:25,500
Make sure to use specific details
from the review.

43
00:02:25,800 --> 00:02:30,000
Write in a concise and professional tone
and sign the email as a customer agent.

44
00:02:30,300 --> 00:02:34,500
And when you're using a language model
to generate text that you're going to show

45
00:02:34,500 --> 00:02:38,666
to a user, it's very important
to have this kind of transparency

46
00:02:38,733 --> 00:02:44,033
and let the user know that
the text they're seeing was generated by,

47
00:02:44,033 --> 00:02:47,166
and then we'll just input the customer
review and the review sentiment.

48
00:02:47,366 --> 00:02:50,766
And also note
that this isn't necessarily important

49
00:02:50,766 --> 00:02:54,700
because we could actually use this prompt
to also extract the review sentiment.

50
00:02:54,700 --> 00:02:57,033
And then in a follow up
step, write the email.

51
00:02:57,033 --> 00:02:59,000
But just for the sake of the example,
well,

52
00:02:59,000 --> 00:03:01,166
we've already extracted the
sentiment from the review.

53
00:03:02,366 --> 00:03:03,333
And so

54
00:03:03,333 --> 00:03:07,200
here we have a response
to the customer addresses,

55
00:03:07,200 --> 00:03:10,233
details that the customer mentioned
in their review

56
00:03:10,966 --> 00:03:15,233
and kind of, as we instructed suggest
that they reach out to customer service

57
00:03:15,233 --> 00:03:20,866
because this is just an air customer
service agent.

58
00:03:20,866 --> 00:03:25,233
Next, we're going to use a parameter
of the language

59
00:03:25,233 --> 00:03:28,300
model called temperature
that will allow us to

60
00:03:29,566 --> 00:03:32,433
change the kind of variety
of the model's responses

61
00:03:32,700 --> 00:03:34,466
so you can kind of think of temperature

62
00:03:34,466 --> 00:03:39,600
as the degree of expiration
or kind of randomness of the model.

63
00:03:39,600 --> 00:03:43,400
And so for this particular phrase,
my favorite food is

64
00:03:43,800 --> 00:03:47,833
the kind of most likely next word
that the model predicts is pizza

65
00:03:48,000 --> 00:03:51,800
and the kind of next to most likely
suggests are sushi and tacos.

66
00:03:51,966 --> 00:03:55,033
And so at a temperature of zero,
the model will always choose

67
00:03:55,033 --> 00:03:57,466
the most likely next word,
which in this case is pizza.

68
00:03:57,633 --> 00:04:01,800
And at a higher temperature, it will also
choose one of the less likely words

69
00:04:01,800 --> 00:04:05,400
and even at an even higher temperature,
it might even choose tacos,

70
00:04:05,400 --> 00:04:09,400
which only kind of has
a 5% chance of being chosen.

71
00:04:10,533 --> 00:04:11,700
And you can imagine that kind

72
00:04:11,700 --> 00:04:15,600
of as the model continues,
this final response.

73
00:04:15,600 --> 00:04:19,233
So my favorite food is pizza, and it kind
of continues to generate more words.

74
00:04:19,466 --> 00:04:22,433
This response will kind of diverge
from the response.

75
00:04:22,800 --> 00:04:25,166
The first response,
which is my favorite food, is tacos.

76
00:04:25,300 --> 00:04:27,166
And so as a kind of model continues,

77
00:04:27,166 --> 00:04:29,466
these two responses will become more
and more different.

78
00:04:30,566 --> 00:04:33,500
In general, when building applications
where you want

79
00:04:33,500 --> 00:04:38,000
a predictable response,
I would recommend using temperature zero.

80
00:04:38,000 --> 00:04:41,433
Throughout all of these videos,
we've been using temperature zero.

81
00:04:41,466 --> 00:04:43,733
And I think that if you're trying
to build a system

82
00:04:43,733 --> 00:04:47,033
that is reliable and predictable,
you should go with this.

83
00:04:47,333 --> 00:04:50,600
If you're trying to use the model
in a more creative way

84
00:04:50,600 --> 00:04:54,166
where you might want a kind of wider
variety of different outputs,

85
00:04:54,366 --> 00:04:56,466
you might want to use
a higher temperature.

86
00:04:57,366 --> 00:05:01,433
So now let's take this same prompt
that we just used

87
00:05:01,866 --> 00:05:05,933
and let's try generating an email,
but let's use a higher temperature.

88
00:05:06,433 --> 00:05:09,000
So in our get completion function

89
00:05:09,500 --> 00:05:11,933
that we've been using
throughout the videos,

90
00:05:11,933 --> 00:05:15,166
we have kind of specified a model
and then also a temperature,

91
00:05:15,166 --> 00:05:16,566
but we've kind of set them to default.

92
00:05:16,566 --> 00:05:21,233
So now let's try burying the temperature.

93
00:05:21,233 --> 00:05:23,600
So we use the prompt and then

94
00:05:27,533 --> 00:05:27,966
let's try a

95
00:05:27,966 --> 00:05:33,666
temperature 0.7.

96
00:05:33,666 --> 00:05:37,333
And so with temperature zero,
every time you execute the same prompt,

97
00:05:37,333 --> 00:05:40,400
you should expect the same completion.

98
00:05:40,633 --> 00:05:44,633
Whereas with the temperature 0.7, you will
get a different output every time.

99
00:05:45,133 --> 00:05:47,766
So here we have

100
00:05:47,766 --> 00:05:49,900
email and as you can see, it's different

101
00:05:49,900 --> 00:05:54,600
to the email that we kind of received
previously and let's just execute it again

102
00:05:54,966 --> 00:05:57,533
to show
that we'll get a different email again.

103
00:05:59,066 --> 00:06:01,200
And here we have another different email.

104
00:06:01,700 --> 00:06:06,300
So I recommend that you kind of play
around with temperature yourself.

105
00:06:06,666 --> 00:06:10,566
Maybe you could pause the video now
and try this prompt

106
00:06:10,566 --> 00:06:14,066
with a variety of different temperatures
just to see how the outputs vary.

107
00:06:15,566 --> 00:06:17,200
So to summarize

108
00:06:17,200 --> 00:06:21,300
at higher temperatures, the outputs from
the model are kind of more random.

109
00:06:21,466 --> 00:06:24,233
You can almost think of it
as at higher temperatures.

110
00:06:24,233 --> 00:06:27,833
The assistant is more distractible
but maybe more creative.

111
00:06:29,433 --> 00:06:29,933
In the next

112
00:06:29,933 --> 00:06:34,200
video, we're going to talk more
about the chart completions and format

113
00:06:34,300 --> 00:06:37,466
and how you can create a custom
chat bot using this format.