1
00:00:00,090 --> 00:00:03,960
-: So interestingly enough, just as an example that ChatGPT

2
00:00:03,960 --> 00:00:04,922
isn't always going to give you

3
00:00:04,922 --> 00:00:06,600
the most accurate information,

4
00:00:06,600 --> 00:00:09,030
I was actually able to get it to report something

5
00:00:09,030 --> 00:00:11,370
that is seemingly inconsistent

6
00:00:11,370 --> 00:00:14,670
with what we just saw on the homepage of ChatGPT,

7
00:00:14,670 --> 00:00:17,430
which is that under their methods section,

8
00:00:17,430 --> 00:00:19,350
OpenAI has written that they train the model

9
00:00:19,350 --> 00:00:23,070
using reinforcement learning from human feedback, or RLHF.

10
00:00:23,070 --> 00:00:26,400
So, I am trying to fact-check that with ChatGPT itself.

11
00:00:26,400 --> 00:00:30,900
And queried if he was trained or it was trained using RLHF.

12
00:00:30,900 --> 00:00:33,450
And it responded saying it's not familiar with that term.

13
00:00:33,450 --> 00:00:38,070
So, then I clarify, and sure enough it reports a no.

14
00:00:38,070 --> 00:00:43,070
So, dangerous, sometimes it can sort of lie to you,

15
00:00:43,140 --> 00:00:46,680
but it's never in its intention to lie to you.

16
00:00:46,680 --> 00:00:49,440
It's just because the data that it was trained on

17
00:00:49,440 --> 00:00:52,830
just did not agree with maybe something

18
00:00:52,830 --> 00:00:54,870
that was seemingly updated in the future.

19
00:00:54,870 --> 00:00:56,910
So, this probably may have come after,

20
00:00:56,910 --> 00:00:59,220
or maybe it wasn't included in its training data,

21
00:00:59,220 --> 00:01:01,200
which is a pretty big limitation.

22
00:01:01,200 --> 00:01:02,940
But for a lot of the use cases,

23
00:01:02,940 --> 00:01:05,280
we're gonna be exploring in this course, it's again,

24
00:01:05,280 --> 00:01:07,413
not going to be that much of a hindrance.