WEBVTT

1
00:00:00.000 --> 00:00:01.370
Welcome to the workbench.

2
00:00:01.753 --> 00:00:05.840
In this lesson we will use an assistant
for a small engineering task and decide

3
00:00:05.875 --> 00:00:09.416
what evidence its answer needs.
You do not need a board on the desk.

4
00:00:09.741 --> 00:00:12.864
You need a short piece of source
code, a place to keep notes,

5
00:00:12.946 --> 00:00:16.823
and enough curiosity to ask what
each statement actually establishes.

6
00:00:17.148 --> 00:00:19.122
Our result will be an evidence worksheet

7
00:00:19.157 --> 00:00:23.360
that you can use throughout the course.
The running project is a sensor monitor.

8
00:00:23.604 --> 00:00:25.334
We will explain it, change it,

9
00:00:25.450 --> 00:00:28.689
test its software behavior,
and preserve a useful history.

10
00:00:28.933 --> 00:00:32.462
AI helps with those development
activities. The finished firmware

11
00:00:32.509 --> 00:00:35.875
does not need to contain a language
model or call a cloud service.

12
00:00:36.201 --> 00:00:39.788
Keeping development assistance
separate from AI running on a device

13
00:00:39.881 --> 00:00:43.190
makes this first project easier
to understand and verify.

14
00:00:43.700 --> 00:00:46.231
Look at the short reporting
function in the lesson.

15
00:00:46.614 --> 00:00:49.111
Its input is a floating point temperature.

16
00:00:49.552 --> 00:00:53.882
The comparison asks whether that input
is greater than or equal to twenty-eight.

17
00:00:54.323 --> 00:00:56.912
If so, the function returns the word WARN.

18
00:00:57.191 --> 00:01:00.767
Otherwise it returns OK. The
function is deliberately small.

19
00:01:01.046 --> 00:01:04.587
There is no sensor read, no
timer, no pin configuration,

20
00:01:04.680 --> 00:01:08.581
and no communication with a board in
the excerpt. Before asking for help,

21
00:01:08.650 --> 00:01:13.097
identify the task you can judge. We want
an explanation of this supplied function.

22
00:01:13.480 --> 00:01:17.404
We do not want the assistant to invent
an entire acquisition system around it.

23
00:01:17.787 --> 00:01:21.363
The prompt asks it to separate
statements supported directly by the code

24
00:01:21.398 --> 00:01:24.255
from assumptions needing a
document or a physical test.

25
00:01:24.696 --> 00:01:27.412
This gives the answer a
useful shape without demanding

26
00:01:27.459 --> 00:01:30.536
that it use exactly the same
words as the instructor.

27
00:01:30.977 --> 00:01:34.936
An actual model response may differ
from the saved instructional example.

28
00:01:35.377 --> 00:01:39.312
Preserve what your session returns.
If your account is unavailable,

29
00:01:39.521 --> 00:01:42.482
use the chapter's clearly
labeled illustrative response.

30
00:01:43.086 --> 00:01:45.942
You can practice evaluating
claims without pretending

31
00:01:45.988 --> 00:01:48.914
that a new conversation
occurred. What matters here

32
00:01:48.972 --> 00:01:52.234
is whether you can connect each
claim with an appropriate check.

33
00:01:52.931 --> 00:01:56.333
Start with the statement that
twenty-eight produces WARN.

34
00:01:57.145 --> 00:02:01.255
For an ordinary finite input, the
comparison supports that statement.

35
00:02:01.697 --> 00:02:05.365
Equality is included by the
greater-than-or-equal operator.

36
00:02:05.690 --> 00:02:08.929
We can make the boundary visible
with three example values:

37
00:02:09.173 --> 00:02:13.422
twenty-seven point nine, twenty-eight
point zero, and twenty-eight point one.

38
00:02:13.806 --> 00:02:16.429
Those examples separate
an inclusive threshold

39
00:02:16.511 --> 00:02:18.635
from a strict greater-than comparison.

40
00:02:19.018 --> 00:02:23.825
Now consider the statement that the
function samples a BME280 once a second.

41
00:02:24.151 --> 00:02:28.446
The excerpt contains no sampling call
or scheduling code. The statement might

42
00:02:28.493 --> 00:02:32.904
describe some larger program, but it is
not supported by the material we supplied.

43
00:02:33.346 --> 00:02:37.212
In the worksheet, mark it unresolved
and identify the next check:

44
00:02:37.491 --> 00:02:40.091
inspect the caller and
the actual timing code.

45
00:02:40.335 --> 00:02:44.538
Do not add a timer to the function merely
to make an explanation become true.

46
00:02:44.921 --> 00:02:47.197
The same discipline
applies to a pin number.

47
00:02:47.475 --> 00:02:51.399
An assistant may recognize a board
family and suggest a familiar mapping.

48
00:02:51.724 --> 00:02:54.766
Before relying on that number,
identify the exact variant

49
00:02:54.812 --> 00:02:57.901
and consult its manufacturer
pinout or schematic.

50
00:02:58.284 --> 00:03:02.730
A similarly named board can differ
in a consequential way. Our worksheet

51
00:03:02.777 --> 00:03:06.899
records the source needed and leaves the
claim open until that source is inspected.

52
00:03:07.711 --> 00:03:11.090
A physical accuracy claim needs
a different kind of evidence.

53
00:03:11.415 --> 00:03:15.885
A manufacturer specification can describe
performance under stated conditions.

54
00:03:16.326 --> 00:03:19.472
It cannot show that your particular
assembly is wired correctly

55
00:03:19.913 --> 00:03:23.489
or that your measurement matches a
reference. If you choose the optional

56
00:03:23.535 --> 00:03:27.343
hardware lab later, you will define
and record appropriate observations.

57
00:03:27.854 --> 00:03:31.836
For the core course, software tests
and documented constraints are enough,

58
00:03:32.080 --> 00:03:34.297
provided we describe
their limits honestly.

59
00:03:35.249 --> 00:03:38.651
Notice another small distinction in
the code. The parameter name says

60
00:03:38.825 --> 00:03:42.471
temperature underscore c.
That tells us the intended unit

61
00:03:42.529 --> 00:03:46.453
is Celsius. The function itself cannot
inspect the meaning of a number.

62
00:03:47.150 --> 00:03:50.528
If a caller passes Fahrenheit,
the comparison still runs.

63
00:03:50.969 --> 00:03:53.778
To support the complete claim
that the input is Celsius,

64
00:03:53.953 --> 00:03:56.518
inspect the caller contract
and conversion path.

65
00:03:56.727 --> 00:04:00.907
A useful code explanation can reveal
exactly where more context is needed.

66
00:04:01.232 --> 00:04:04.761
Open the evidence worksheet and make
one row per consequential claim.

67
00:04:05.365 --> 00:04:09.997
Use columns for the claim, current
evidence label, supporting line or source,

68
00:04:10.322 --> 00:04:14.177
next check, and result.
Keep the next check concrete.

69
00:04:14.560 --> 00:04:17.381
Write inspect the caller for
its scheduling condition,

70
00:04:17.660 --> 00:04:21.062
rather than verify later.
Write compare three boundary

71
00:04:21.120 --> 00:04:24.928
values against the requirement, rather
than make sure the program works.

72
00:04:25.740 --> 00:04:28.608
There are several kinds of
assistant around the workbench.

73
00:04:29.119 --> 00:04:34.111
A model generates output. An application
provides the chat or editor interface.

74
00:04:34.552 --> 00:04:38.407
An agent can use tools, depending on
the permissions and features available.

75
00:04:39.011 --> 00:04:43.585
A command printed in an answer is
proposed text until something executes it.

76
00:04:44.189 --> 00:04:47.776
Later lessons will show how to
inspect tool actions, command results,

77
00:04:47.834 --> 00:04:51.329
and changed files instead of
relying only on a final summary.

78
00:04:51.770 --> 00:04:55.195
For your exercise, classify
three claims and rewrite

79
00:04:55.219 --> 00:04:58.446
the broad conclusion that the
device is ready to deploy.

80
00:04:58.887 --> 00:05:01.790
A better conclusion identifies
the positive result:

81
00:05:02.115 --> 00:05:05.969
the shown function implements the
stated finite threshold comparison.

82
00:05:06.410 --> 00:05:09.417
It also names the remaining
questions: acquisition,

83
00:05:09.626 --> 00:05:12.610
invalid values, timing,
and physical operation.

84
00:05:13.214 --> 00:05:15.907
You are preserving what the
evidence supports while making

85
00:05:15.954 --> 00:05:19.355
the next work visible. Check
your worksheet before continuing.

86
00:05:19.739 --> 00:05:21.910
Can another person see which line supports

87
00:05:21.945 --> 00:05:25.022
the comparison? Can they tell
which reference is missing?

88
00:05:25.405 --> 00:05:29.073
Can they distinguish a synthetic
test from a physical observation?

89
00:05:29.515 --> 00:05:32.777
If so, you have created a
useful engineering artifact.

90
00:05:33.218 --> 00:05:36.411
The goal is not to make the
assistant sound less confident.

91
00:05:36.794 --> 00:05:39.534
The goal is to make your
own decision traceable.

92
00:05:39.918 --> 00:05:44.155
Consider how you would record a check that
has not happened yet. Leave its result

93
00:05:44.236 --> 00:05:48.068
blank or write pending, and keep
the proposed method beside it.

94
00:05:48.451 --> 00:05:51.087
That small distinction
prevents a plan from becoming

95
00:05:51.133 --> 00:05:53.769
a success claim when
you return days later.

96
00:05:54.152 --> 00:05:59.295
When a real check occurs, add the date,
input, and actual result. The worksheet

97
00:05:59.330 --> 00:06:03.277
then becomes a record of learning rather
than a list of reassuring predictions.

98
00:06:03.718 --> 00:06:07.828
Finish by choosing one task from
a project you already understand.

99
00:06:08.270 --> 00:06:11.288
Name the material you can
provide, the result you want,

100
00:06:11.346 --> 00:06:14.167
and the check you would
accept. Keep that short note.

101
00:06:14.411 --> 00:06:17.279
We will expand it into a
complete engineering brief after

102
00:06:17.349 --> 00:06:20.042
exploring why context
helps a language model

103
00:06:20.367 --> 00:06:23.409
and why context alone never
guarantees correctness.