WEBVTT

1
00:00:00.000 --> 00:00:07.100
This overview explains why input width alone
is not enough to choose a fast execution path.

2
00:00:07.100 --> 00:00:15.263
This matters because a selector can look good on typical cases
while making one case much slower than the fixed control.

3
00:00:15.263 --> 00:00:15.267
This matters because a selector can look good on typical cases
while making one case much slower than the fixed control.

4
00:00:15.567 --> 00:00:19.133
Both rules mention four Boolean inputs.

5
00:00:19.133 --> 00:00:24.833
The second repeats A across several branches and offers sharing or factoring opportunities

6
00:00:24.833 --> 00:00:27.900
that the first does not present in the same way.

7
00:00:27.900 --> 00:00:34.467
Width is one descriptor of the job, not a complete description of its structure.

8
00:00:34.467 --> 00:00:38.200
Imagine a router whose feature extraction costs three units and whose

9
00:00:38.200 --> 00:00:41.100
chosen exact arm costs nine.

10
00:00:41.100 --> 00:00:44.067
A fixed exact method costs ten.

11
00:00:44.067 --> 00:00:49.533
The routed total is twelve, so the selected arm's attractive nine-unit execution does not

12
00:00:49.533 --> 00:00:52.500
make the dispatcher a winner.

13
00:00:52.500 --> 00:00:57.967
With regret defined as routed total divided by the fixed reference, this case has a ratio

14
00:00:57.967 --> 00:00:59.733
of one point two.

15
00:00:59.733 --> 00:01:03.667
Keep feature extraction and any conversion inside the boundary when the

16
00:01:03.667 --> 00:01:05.400
caller must pay them.

17
00:01:05.400 --> 00:01:09.433
The ratio convention must be declared.

18
00:01:09.433 --> 00:01:09.435
A selector chooses an execution path
using information available before it runs.

19
00:01:09.435 --> 00:01:15.833
A selector chooses an execution path
using information available before it runs.

20
00:01:15.833 --> 00:01:19.700
Count the cost of obtaining that information.

21
00:01:19.700 --> 00:01:25.833
Here, regret is the selector's total time divided by the fixed control's time.

22
00:01:25.833 --> 00:01:29.267
Every case must stay below two.

23
00:01:29.267 --> 00:01:33.767
A good median alone can hide a severe losing case.

24
00:01:33.767 --> 00:01:39.967
Imagine a toy policy that chooses the
packed-word engine above a width threshold.

25
00:01:39.967 --> 00:01:46.300
Most cases cost one unit relative to the fixed control, but one costs ten.

26
00:01:46.300 --> 00:01:52.100
The median looks harmless while that
case takes ten times the control's time.

27
00:01:52.100 --> 00:01:58.967
If the frozen rule requires every regret ratio below two, this policy fails.

28
00:01:58.967 --> 00:02:05.300
This arithmetic illustrates a gate; it is not a new retained selector experiment.

29
00:02:05.300 --> 00:02:12.933
The corrected focused width study did not establish
a universal switch rule and reused validation.

30
00:02:12.933 --> 00:02:19.000
Later retained q64 selector work also promoted no learned route.

31
00:02:19.000 --> 00:02:21.000
Keep those cohorts distinct,

32
00:02:21.000 --> 00:02:29.060
and retain the exact fixed fallback rather than silently
installing a heuristic suggested by the animation.

33
00:02:29.060 --> 00:02:29.067
and retain the exact fixed fallback rather than silently
installing a heuristic suggested by the animation.

34
00:02:29.367 --> 00:02:32.433
Sort these four illustrative ratios.

35
00:02:32.433 --> 00:02:38.067
The median is the average of zero point nine and one, or zero point nine five.

36
00:02:38.067 --> 00:02:44.000
The typical-looking summary can coexist with a tenfold bad case.

37
00:02:44.000 --> 00:02:48.500
If the declared acceptance rule requires every case to stay below two,

38
00:02:48.500 --> 00:02:51.200
the tenfold case fails it.

39
00:02:51.200 --> 00:02:54.467
The good median cannot override that requirement.

40
00:02:54.467 --> 00:02:59.867
This is a direct conclusion from the stated gate, not a quiz for the viewer.

41
00:02:59.867 --> 00:03:04.500
A selector revised after observing the bad case is a new candidate.

42
00:03:04.500 --> 00:03:09.367
It needs an appropriately fresh confirmation set under a frozen protocol.

43
00:03:09.367 --> 00:03:13.467
Retesting on the same exposed cases can demonstrate repair there but

44
00:03:13.467 --> 00:03:18.133
does not establish unseen-case transfer.

45
00:03:18.133 --> 00:03:18.147
In the toy example, case four takes ten times the
control's time, so the policy fails the below-two rule.

46
00:03:18.147 --> 00:03:26.467
In the toy example, case four takes ten times the
control's time, so the policy fails the below-two rule.

47
00:03:26.467 --> 00:03:36.059
The takeaway is to charge the full selection cost and keep
the fixed fallback until a frozen policy passes its validation.