WEBVTT

1
00:00:00.000 --> 00:00:04.000
Both choose from answers supplied with your question

2
00:00:04.000 --> 00:00:13.000
Jev returns a selected answer and its probabilities

3
00:00:13.000 --> 00:00:24.000
Venn scores answers using numeric representations of meaning

4
00:00:24.000 --> 00:00:28.000
Now predict outcomes from complaints to this company

5
00:00:28.000 --> 00:00:34.000
Use Jev's label to group past outcomes

6
00:00:34.000 --> 00:00:42.000
Venn retrieves similar complaints with recorded outcomes

7
00:00:42.000 --> 00:00:48.000
Evaluate outcome rankings on previously unseen complaints

8
00:00:48.000 --> 00:00:52.000
Now compare only complaints requesting a refund

9
00:00:52.000 --> 00:00:57.000
Imagine removing the complaint archive

10
00:00:57.000 --> 00:01:04.000
Now learn industry labels from company filings

11
00:01:04.000 --> 00:01:09.000
Test both methods on the cookbook's filings

12
00:01:09.000 --> 00:01:15.000
Then test on additional held-out company filings

13
00:01:15.000 --> 00:01:21.000
Inspect one filing's predictions

14
00:01:21.000 --> 00:01:29.000
Venn learns industry labels from past filings

15
00:01:29.000 --> 00:01:34.000
Predict the critic's score from wine tasting notes

16
00:01:34.000 --> 00:01:42.000
Jev builds features by answering questions about notes

17
00:01:42.000 --> 00:01:56.000
Venn fits a prediction head on stored embeddings

18
00:01:56.000 --> 00:02:02.000
Evaluate critic scores on previously unseen wine notes

19
00:02:02.000 --> 00:02:07.000
Both learn from the same 1,200 training notes

20
00:02:07.000 --> 00:02:14.000
Price uses a different, already-stored embedding cache

21
00:02:14.000 --> 00:02:20.000
Venn's time excludes the already-completed embedding step

22
00:02:20.000 --> 00:02:23.000
Local predictions require no additional API charges

23
00:02:23.000 --> 00:02:27.000
Additional tests favor Jev on these reading tasks

24
00:02:27.000 --> 00:02:30.000
Build decisions from your own data