WEBVTT

1
00:00:00.000 --> 00:00:03.971
A language model can be helpful without
being a database of verified project

2
00:00:04.029 --> 00:00:08.557
facts. This lesson gives us a practical
mental model for that difference.

3
00:00:09.068 --> 00:00:12.679
We will compare answers with and
without a short source passage,

4
00:00:13.004 --> 00:00:15.836
then ask what the
comparison actually proves.

5
00:00:16.220 --> 00:00:20.086
You do not need to implement a transformer
or inspect hidden reasoning to use

6
00:00:20.120 --> 00:00:23.186
the result at your workbench.
Begin with tokens.

7
00:00:23.464 --> 00:00:26.738
A token is a unit used to
represent input or output.

8
00:00:27.064 --> 00:00:31.150
It may be a word, part of a word,
punctuation, or another encoded unit.

9
00:00:31.533 --> 00:00:34.471
Do not assume that one token
means one English word.

10
00:00:34.796 --> 00:00:37.582
A long component identifier,
a block of code,

11
00:00:37.617 --> 00:00:40.183
and an ordinary sentence
can divide differently.

12
00:00:40.787 --> 00:00:43.086
Token limits therefore
cannot be understood

13
00:00:43.155 --> 00:00:45.524
simply by counting lines in an editor.

14
00:00:46.221 --> 00:00:50.493
A useful simplified picture is that
the model predicts a next token

15
00:00:50.551 --> 00:00:55.416
from the available context, adds it to
the growing sequence, and continues.

16
00:00:56.019 --> 00:01:00.315
Learned patterns can support explanations,
summaries, and useful code.

17
00:01:00.640 --> 00:01:03.821
That process does not automatically
inspect the current release

18
00:01:03.856 --> 00:01:07.490
of your sensor library or run the
test whose output it describes.

19
00:01:07.931 --> 00:01:11.298
A tool or supplied reference
is needed for those actions.

20
00:01:11.623 --> 00:01:14.908
The conceptual diagram separates
training from inference.

21
00:01:15.349 --> 00:01:20.016
Training adjusts model parameters
using data and an optimization process.

22
00:01:20.458 --> 00:01:23.778
Further training can shape how
a model follows instructions.

23
00:01:24.161 --> 00:01:27.354
Inference is using the resulting
model for a new request.

24
00:01:27.737 --> 00:01:30.395
When you paste a project
note into a conversation,

25
00:01:30.442 --> 00:01:33.344
you are supplying information
for the current interaction.

26
00:01:33.553 --> 00:01:36.978
You have not performed the same
operation as retraining the model.

27
00:01:37.303 --> 00:01:39.753
That distinction matters
when you resume work.

28
00:01:40.078 --> 00:01:42.261
If you correct a requirement in one chat,

29
00:01:42.412 --> 00:01:45.291
a later session may need the
corrected document again.

30
00:01:45.616 --> 00:01:49.808
Save the requirement in a project
file with a useful name. Do not rely

31
00:01:49.877 --> 00:01:54.057
on the assistant remembering a correction
forever because it agreed with you once.

32
00:01:54.382 --> 00:01:58.933
Later modules will make those persistent
documents part of a repeatable workflow.

33
00:01:59.258 --> 00:02:02.846
The second diagram shows context
entering the interaction.

34
00:02:03.229 --> 00:02:05.853
The request, selected reference material,

35
00:02:05.969 --> 00:02:10.357
retained conversation information,
and tool results can all contribute.

36
00:02:10.741 --> 00:02:13.260
The diagram is a teaching simplification.

37
00:02:13.643 --> 00:02:17.195
It is not a private trace of
the model's internal process.

38
00:02:17.799 --> 00:02:21.421
Our practical influence is over
the quality and organization

39
00:02:21.468 --> 00:02:25.206
of information that we supply and
the checks we perform afterward.

40
00:02:25.810 --> 00:02:28.736
Attention is a mechanism that
allows relationships among

41
00:02:28.828 --> 00:02:31.568
input representations
to contribute to output.

42
00:02:31.893 --> 00:02:35.214
That helps explain why a short note
can connect a temperature unit,

43
00:02:35.388 --> 00:02:38.128
threshold, and validity
rule in one answer.

44
00:02:38.453 --> 00:02:41.994
It does not mean every relevant
sentence is always used correctly.

45
00:02:42.319 --> 00:02:45.210
A larger input containing
old and contradictory

46
00:02:45.291 --> 00:02:48.646
rules may make the task
harder rather than easier.

47
00:02:49.029 --> 00:02:51.827
Now open the original BenchNote passage.

48
00:02:52.210 --> 00:02:56.158
This is a fictional teaching
specification written for the exercise,

49
00:02:56.332 --> 00:03:00.036
not a manufacturer's manual.
It says samples are synthetic,

50
00:03:00.210 --> 00:03:03.925
temperatures use Celsius,
invalid samples produce FAULT,

51
00:03:03.995 --> 00:03:07.675
and valid samples at or above
twenty-eight produce WARN.

52
00:03:07.849 --> 00:03:12.749
It explicitly describes a simplified
policy without the capstone's hysteresis.

53
00:03:13.074 --> 00:03:14.932
We will keep those boundaries visible.

54
00:03:15.175 --> 00:03:19.622
Before supplying the note, ask how
BenchNote handles an invalid sample.

55
00:03:19.901 --> 00:03:21.538
Save the actual answer.

56
00:03:21.781 --> 00:03:25.253
The assistant may say it lacks
enough information. That is a good

57
00:03:25.311 --> 00:03:28.829
response because the product-specific
rule has not been supplied.

58
00:03:29.107 --> 00:03:33.763
It may instead suggest a typical behavior,
such as retaining the previous reading.

59
00:03:34.205 --> 00:03:37.049
If it presents that suggestion
as a fact about BenchNote,

60
00:03:37.258 --> 00:03:39.278
mark the claim as unsupported.

61
00:03:39.789 --> 00:03:42.796
Now provide the complete passage
and ask the same question.

62
00:03:43.307 --> 00:03:46.349
Add a clear instruction to use
this note for product-specific

63
00:03:46.407 --> 00:03:48.601
claims and identify missing facts.

64
00:03:49.043 --> 00:03:52.189
The answer should say that
invalid input produces FAULT

65
00:03:52.514 --> 00:03:56.368
and that its numeric temperature
does not select OK or WARN.

66
00:03:56.752 --> 00:03:59.201
Check that statement
against the actual sentence.

67
00:03:59.480 --> 00:04:02.290
A correct answer is useful
because you can trace it,

68
00:04:02.615 --> 00:04:06.504
not merely because it sounds more
detailed. Repeat the with-reference

69
00:04:06.539 --> 00:04:11.183
request in a fresh conversation or compare
the supplied original alternatives.

70
00:04:11.694 --> 00:04:14.910
Wording may differ. One answer
may discuss the validity

71
00:04:14.980 --> 00:04:17.929
flag first and another
may name the state first.

72
00:04:18.312 --> 00:04:21.795
Those differences are not
necessarily technical disagreements.

73
00:04:22.178 --> 00:04:25.777
Your worksheet should compare
claims: fault behavior, unit,

74
00:04:25.870 --> 00:04:29.910
valid threshold boundary, and
information that remains unspecified.

75
00:04:30.351 --> 00:04:34.148
Ask how many consecutive valid
samples are required to recover.

76
00:04:34.658 --> 00:04:37.387
This introductory note
does not define that count.

77
00:04:37.631 --> 00:04:40.591
The correct judgment is that
another requirement is needed.

78
00:04:40.870 --> 00:04:44.794
A plausible number is still unsupported.
Keep the gap visible instead

79
00:04:44.829 --> 00:04:48.776
of filling it with a pattern the
assistant recognizes from another project.

80
00:04:49.287 --> 00:04:52.933
The final sensor monitor will have
its own explicit recovery requirement

81
00:04:53.060 --> 00:04:57.008
later in the course. What does a
successful comparison establish?

82
00:04:57.518 --> 00:05:01.396
It shows that these answers handled
this supplied information appropriately.

83
00:05:01.675 --> 00:05:04.798
It does not show that every
future answer will be correct.

84
00:05:05.123 --> 00:05:09.686
A repeated error can remain consistent,
and different wording can remain correct.

85
00:05:10.069 --> 00:05:14.643
Avoid treating agreement among responses
as a substitute for a source or a test.

86
00:05:14.968 --> 00:05:16.803
If an answer contradicts the note,

87
00:05:17.081 --> 00:05:20.007
isolate the claim and point
to the governing sentence.

88
00:05:20.518 --> 00:05:24.152
Ask for a corrected artifact,
then inspect it yourself.

89
00:05:24.593 --> 00:05:27.809
You can also save the proposed
alternative as a design

90
00:05:27.867 --> 00:05:30.862
idea, clearly separate from
the current requirement.

91
00:05:31.245 --> 00:05:35.483
Do not silently rewrite the requirement
to match a convincing explanation.

92
00:05:36.435 --> 00:05:40.533
There is another useful comparison you
can make without another model call.

93
00:05:41.230 --> 00:05:44.504
Remove one essential sentence
from your own view of the passage

94
00:05:44.678 --> 00:05:47.870
and ask what conclusion you
can still support as a reader.

95
00:05:48.567 --> 00:05:50.355
If the validity rule is missing,

96
00:05:50.529 --> 00:05:53.722
you cannot derive FAULT merely
from the temperature value.

97
00:05:54.233 --> 00:05:58.865
This exercise shows why the information
packet matters independently of any

98
00:05:58.923 --> 00:06:03.323
particular assistant. The brief should
contain the facts needed for the decision,

99
00:06:03.567 --> 00:06:06.876
while the result should preserve
gaps that the packet leaves open.

100
00:06:07.317 --> 00:06:10.429
Complete the comparison
worksheet before advancing.

101
00:06:10.754 --> 00:06:15.572
Keep the original question, the passage,
the actual or labeled sample answers,

102
00:06:15.688 --> 00:06:19.404
and your conclusions together.
Your success check is simple:

103
00:06:19.682 --> 00:06:24.071
another maker can see what information
changed, which claims became grounded,

104
00:06:24.280 --> 00:06:28.401
and which detail is still unknown.
That is a more durable result than

105
00:06:28.483 --> 00:06:32.047
finding a prompt that happened to
produce an impressive paragraph once.