prompt_eng_01_intro_v3.srt

1
00:00:05,105 --> 00:00:05,638
Welcome to

2
00:00:05,638 --> 00:00:08,975
this course on
catch up from engineering for developers

3
00:00:09,376 --> 00:00:14,147
are thrilled to have with me
is a full fit to teach this along with me

4
00:00:14,514 --> 00:00:19,252
She's a member of the technical staff
of Openai and had built the popular chip

5
00:00:19,486 --> 00:00:23,423
retrieval plugin, and a large
part of the work has been teaching people

6
00:00:23,656 --> 00:00:27,360
how to use l M or large language
model technology in products.

7
00:00:27,727 --> 00:00:31,464
She's also a contributor to the Open,
a cookbook that teaches people prompting.

8
00:00:31,464 --> 00:00:34,667
So thrilled to have you with you
and I'm thrilled to be here and share

9
00:00:34,667 --> 00:00:36,703
some prompting best practices
with you all.

10
00:00:38,204 --> 00:00:41,708
So there's been a lot of material
on the internet for prompting

11
00:00:41,708 --> 00:00:44,944
with articles like 30 prompts
everyone has to do.

12
00:00:45,678 --> 00:00:50,250
A lot of that has been focused
on the chat Jupiter Web user interface,

13
00:00:50,250 --> 00:00:54,421
which many people are using to do specific
and often one off tasks.

14
00:00:54,888 --> 00:00:59,993
But I think the power of our like language
models as a developer to

15
00:01:00,360 --> 00:01:04,798
that is using API calls to OEMs to quickly
build software applications.

16
00:01:04,798 --> 00:01:07,734
I think that is still very
underappreciated.

17
00:01:08,201 --> 00:01:09,602
In fact, my team at A.I.

18
00:01:09,602 --> 00:01:13,239
Fund, which is a sister company
to deep A.I., has been working

19
00:01:13,239 --> 00:01:16,543
with many startups
on applying these technologies

20
00:01:16,743 --> 00:01:20,313
to many different applications,
and it's been exciting to see

21
00:01:20,513 --> 00:01:25,151
what APIs can enable developers
to very quickly build.

22
00:01:25,718 --> 00:01:29,823
So in this course, we'll share with you
some of the possibilities for what

23
00:01:29,889 --> 00:01:33,660
you can do as well as best practices
for how you can do them.

24
00:01:34,761 --> 00:01:37,397
There's a lot of material to cover fast.

25
00:01:37,397 --> 00:01:40,934
You'll learn best, some prompting
best practices for software development

26
00:01:41,134 --> 00:01:45,138
that will cover some common use cases
summarizing, inferring,

27
00:01:45,271 --> 00:01:47,040
transforming, expanding.

28
00:01:47,040 --> 00:01:49,576
And then you'll build a chatbot
using an Al-Alam.

29
00:01:50,310 --> 00:01:51,044
We hope that

30
00:01:51,044 --> 00:01:54,314
this will spark your imagination
about new applications that you can build.

31
00:01:55,148 --> 00:01:59,853
So in the development of large language
models, OEMs have been broadly two types

32
00:01:59,853 --> 00:02:02,555
of of which I'm going to refer to as base

33
00:02:03,456 --> 00:02:06,059
and instruction tunes elements.

34
00:02:06,059 --> 00:02:09,496
So Basil has been trained
to predict the next words

35
00:02:09,829 --> 00:02:13,500
based on text training data, often
trained on large amount of data

36
00:02:13,666 --> 00:02:15,301
from the Internet and other sources

37
00:02:15,301 --> 00:02:19,472
to figure out what's the next
most likely word to follow.

38
00:02:19,472 --> 00:02:21,574
So, for example,
if you were to prompt this,

39
00:02:21,808 --> 00:02:25,378
once upon a time there was a unicorn,
it may complete.

40
00:02:25,378 --> 00:02:27,747
This does is it
may predict the next several words

41
00:02:27,747 --> 00:02:30,550
are that live in the magical forest
of all unicorn friends.

42
00:02:31,818 --> 00:02:32,318
But if you

43
00:02:32,318 --> 00:02:35,822
were to prompt us with
what is the capital of France, then

44
00:02:36,222 --> 00:02:41,694
based on what articles on the internet
might have is quite possible to the base,

45
00:02:41,694 --> 00:02:45,765
our team will complete this
with what is France's largest city,

46
00:02:45,765 --> 00:02:49,402
what is France's population, and so on,
because articles on the internet

47
00:02:49,602 --> 00:02:54,407
could quite plausibly be lists of quiz
questions about the country of France.

48
00:02:55,508 --> 00:02:58,044
In contrast, an instruction tuned out,

49
00:02:58,678 --> 00:03:03,650
which is where a lot of momentum
of research and practice has been going

50
00:03:04,050 --> 00:03:07,921
and instruction
has been trained to follow instructions.

51
00:03:07,921 --> 00:03:11,157
So if you were to ask
what is the capital of France is much

52
00:03:11,157 --> 00:03:15,028
more likely to output
something like two cups of France's Paris.

53
00:03:15,662 --> 00:03:19,132
So the way that instruction tune
owners are typically trained

54
00:03:19,132 --> 00:03:23,736
is you start off with a base, doesn't
train the huge amount of text data

55
00:03:24,003 --> 00:03:28,308
and further train it further
fine, tune it within present outputs

56
00:03:28,341 --> 00:03:32,145
the instructions and good attempts
to follow those instructions

57
00:03:32,512 --> 00:03:36,683
and then often further refined
using a technique called HSF reinforcement

58
00:03:36,683 --> 00:03:40,520
learning from human feedback
to make the system better,

59
00:03:40,520 --> 00:03:44,290
able to be helpful and follow instructions
because instructions

60
00:03:45,091 --> 00:03:48,528
have been trained
to be helpful on this and home this.

61
00:03:48,962 --> 00:03:52,232
So for example,
they're less likely to output problematic

62
00:03:52,365 --> 00:03:55,201
text such as toxic outputs
compared to base alarm.

63
00:03:55,868 --> 00:03:59,872
A lot of the practical usage scenarios
have been shifting to other instructions.

64
00:03:59,872 --> 00:04:03,476
You know, as some of the best practices
you find on the Internet

65
00:04:03,743 --> 00:04:06,012
may be more suited for a base element.

66
00:04:06,279 --> 00:04:10,850
But for most practical applications today,
we would recommend most people instead

67
00:04:10,850 --> 00:04:15,121
focus on instruction to use ohms,
which are easier to use,

68
00:04:15,121 --> 00:04:19,492
and also because of the work of Open the
I and other companies

69
00:04:19,692 --> 00:04:21,728
becoming safer and more aligned.

70
00:04:22,695 --> 00:04:23,263
So this

71
00:04:23,263 --> 00:04:27,400
course will focus on best practices
for instruction to our homes,

72
00:04:28,101 --> 00:04:31,704
which is what we recommend
you use for most of your applications.

73
00:04:32,338 --> 00:04:33,172
Before moving on,

74
00:04:33,172 --> 00:04:37,310
I just want to acknowledge the team
from opening and Deployment II

75
00:04:37,377 --> 00:04:41,614
that had contributed to the materials
that is in I will be presenting.

76
00:04:41,614 --> 00:04:44,817
I'm very grateful to Andrew Main, Joe
Palermo, Boris

77
00:04:44,817 --> 00:04:48,655
Power,
Ted Sanders and Lily Wang from Open Air.

78
00:04:48,688 --> 00:04:51,324
They were very involved
with us, brainstorming materials,

79
00:04:51,324 --> 00:04:55,128
vetting the materials to put together
the curriculum for this short course.

80
00:04:55,328 --> 00:04:57,096
And I'm also grateful for the tip

81
00:04:57,096 --> 00:05:00,900
learning site for the work of Jeff
Ludwig at issue and Tony Nelson.

82
00:05:01,334 --> 00:05:03,603
So when you use an instruction to

83
00:05:05,038 --> 00:05:07,707
think of giving instructions
to another person,

84
00:05:08,041 --> 00:05:11,978
say someone that's smart but doesn't know
the specifics of your task.

85
00:05:12,412 --> 00:05:16,683
So when it doesn't work, sometimes it's
because the instructions

86
00:05:16,683 --> 00:05:17,650
weren't clear enough.

87
00:05:17,650 --> 00:05:22,088
For example, if you were to say, Please
write me something about Alan Turing.

88
00:05:22,555 --> 00:05:26,225
Well, in addition to that,
it can be helpful to be clear

89
00:05:26,392 --> 00:05:30,697
about whether you want the text
to focus on his scientific work

90
00:05:30,697 --> 00:05:34,334
or his personal life
or his role in history or something else.

91
00:05:34,834 --> 00:05:39,972
And if you specify what you want
the tone of the text to be shouldn't

92
00:05:40,006 --> 00:05:43,076
take on a tone like a professional
journalists were right?

93
00:05:43,076 --> 00:05:44,677
Or is it more of a casual note

94
00:05:44,677 --> 00:05:48,648
that you dash off to a friend
that hopes to generate what you want?

95
00:05:49,082 --> 00:05:53,720
And of course, if you picture yourself
asking, see a fresh college graduate

96
00:05:54,053 --> 00:05:57,824
to carry out this task for you,
if you can even specify

97
00:05:57,890 --> 00:06:00,960
what snippets of texts
they should read in advance to write

98
00:06:00,960 --> 00:06:04,564
this text about Alan Turing,
then that even better sets up

99
00:06:04,630 --> 00:06:08,334
that fresh college track for success
to carry out this task for you.

100
00:06:09,001 --> 00:06:14,307
So in the next video you see examples
of how to be clear and specific,

101
00:06:14,307 --> 00:06:17,377
which is an important
principle of prompting ohms.

102
00:06:17,677 --> 00:06:21,547
And you also learned from either
a second principle of prompting

103
00:06:21,814 --> 00:06:27,253
that is giving a little time to think so
that let's go on to the next video.