WEBVTT

NOTE AI Engineering Visualised: https://understandingdata.com/ai-engineering-visualised/parallel-work-synthesis/

1
00:00:00.103 --> 00:00:03.966
Should a coffee subscription add fortnightly delivery?

2
00:00:03.966 --> 00:00:12.479
An orchestrator splits the question into four sub-questions, demand, cost, packing capacity and rivals, one per worker.

3
00:00:12.479 --> 00:00:19.457
Run at the same time, the four workers finish in 2.2 seconds, against 7.2 seconds one after another.

4
00:00:19.457 --> 00:00:22.178
Then the merge alone takes 2.6 seconds.

5
00:00:22.178 --> 00:00:25.109
In all five runs, it was the slowest step.

6
00:00:25.109 --> 00:00:32.160
Give two workers overlapping questions, and both return the same two figures, 38% and 9.1%.

7
00:00:32.160 --> 00:00:38.513
Across five runs, the second never added a new figure, only about 600 more tokens.

8
00:00:38.513 --> 00:00:40.543
Now split by source instead.

9
00:00:40.543 --> 00:00:43.879
The survey worker reads that 38% would switch.

10
00:00:43.879 --> 00:00:46.852
The pilot worker sees that only 9.1% did.

11
00:00:46.852 --> 00:00:49.753
Neither worker can see the disagreement.

12
00:00:49.753 --> 00:00:51.638
Only the merge holds both.

13
00:00:51.638 --> 00:00:54.555
Without a merge rule, the conflict disappears.

14
00:00:54.555 --> 00:01:01.529
Merged by question, the pilot's figure overwrote the survey's, only because it finished 13 milliseconds later.

15
00:01:01.529 --> 00:01:06.348
Handed both figures, the lead never said they disagree, in any of five runs.

16
00:01:06.348 --> 00:01:12.364
Write the rule down: keep every figure with its source, and when two disagree, state both.

17
00:01:12.364 --> 00:01:19.650
Now all five answers named the conflict, and the advice changed from no, to it depends which figure is right.

18
00:01:19.650 --> 00:01:22.007
Parallel time is the slowest branch.

19
00:01:22.007 --> 00:01:28.619
One worker was asked to check all thirty rivals row by row, and everything waited 9.4 seconds for it.

20
00:01:28.619 --> 00:01:34.249
With a six-second deadline, the merge went ahead, and said the rival data was missing.

21
00:01:34.249 --> 00:01:41.974
So split only questions that don't overlap, give every branch a deadline, and write the merge rule before you split.

22
00:01:41.974 --> 00:01:45.370
The merge is the one step no worker can do for you.