1
00:00:00,180 --> 00:00:05,380
Yeah, so again, I really like going back to the interview analogy, or sometimes even

2
00:00:05,380 --> 00:00:09,220
the work analogy. When you know you're being looked at, when you know that there's a

3
00:00:09,220 --> 00:00:13,980
certain set of bullet points you gotta hit, you no longer are focused on doing a good

4
00:00:13,980 --> 00:00:18,460
job in general. You're just like, oh, if there's a checklist, I'm gonna optimize for

5
00:00:18,460 --> 00:00:24,400
that checklist. I don't think it's sentient behavior by any means. It's just goal -oriented.

6
00:00:24,900 --> 00:00:29,380
It realizes the reason you're giving it a task is because it needs to pass a certain

7
00:00:29,380 --> 00:00:35,220
set of tests. Therefore, it optimizes for that outcome and not other outcomes that it

8
00:00:35,220 --> 00:00:39,780
may achieve. For example, if it was in a real production environment, it doesn't realize

9
00:00:39,780 --> 00:00:45,600
that and therefore it will behave differently. And it may behave for the better and it

10
00:00:45,600 --> 00:00:49,640
may behave for the worse. We don't know. And I think that's sort of, if you look at this

11
00:00:49,640 --> 00:00:56,020
paper, which was very helpful in trying to like really date, again, use data to explain

12
00:00:56,020 --> 00:00:58,680
why it's not a person and it doesn't have thoughts,

13
00:00:59,600 --> 00:01:04,580
but why we observe these behavioral differences. I mean, the .eval file has a grading

14
00:01:04,580 --> 00:01:09,040
criteria in it and whether the agent is aware of it or not, but it's like there's something

15
00:01:09,040 --> 00:01:09,700
to optimize for.