WEBVTT

1
00:00:04.480 --> 00:00:06.160
Leonid, there are concerns

2
00:00:06.160 --> 00:00:09.760
from business leaders who are
deploying GenAI in companies.

3
00:00:09.760 --> 00:00:13.280
One of them is privacy
and security of data.

4
00:00:13.280 --> 00:00:15.600
Any guidance on your side?

5
00:00:15.600 --> 00:00:19.560
First of all, definitely privacy and
security are a very important problem,

6
00:00:19.560 --> 00:00:21.960
but it is much more of an engineering problem

7
00:00:21.960 --> 00:00:24.480
than a scientific problem.

8
00:00:24.480 --> 00:00:27.720
So it has to be dealt
with in many ways

9
00:00:27.720 --> 00:00:32.360
similar to how we deal with other
information systems when we use data.

10
00:00:32.360 --> 00:00:34.280
From the point of view of modelling,

11
00:00:34.280 --> 00:00:36.520
there are really two options that businesses have:

12
00:00:36.520 --> 00:00:39.760
either use models that has been pre-trained for them

13
00:00:39.760 --> 00:00:42.520
as a service from model providers,

14
00:00:42.520 --> 00:00:44.360
or build on-premise,

15
00:00:44.360 --> 00:00:47.440
depending on their goals and on the level of security

16
00:00:47.440 --> 00:00:49.440
and privacy they want for their data.

17
00:00:49.440 --> 00:00:54.440
Leonid, are there ways to control the risk of hallucination?

18
00:00:54.560 --> 00:00:56.320
There are several ways.

19
00:00:56.320 --> 00:00:59.520
One way is when you're training the model,

20
00:00:59.520 --> 00:01:01.240
and if you're training the model yourself,

21
00:01:01.240 --> 00:01:03.760
you can actually limit the amount of data

22
00:01:03.760 --> 00:01:05.920
you use and the quality of the data.

23
00:01:05.920 --> 00:01:07.960
Of course, the model will be smaller than

24
00:01:07.960 --> 00:01:11.760
what the model provided to you by, say, OpenAI.

25
00:01:11.760 --> 00:01:15.640
But then there are also ways to control hallucination

26
00:01:15.640 --> 00:01:18.240
during the execution time,
during inference time.

27
00:01:18.240 --> 00:01:23.000
For example, you can provide very specific prompts to the model.

28
00:01:23.000 --> 00:01:26.760
You can actually ask the model to self-reflect on the results,

29
00:01:26.760 --> 00:01:30.240
or you can, for example, run an ensemble of models

30
00:01:30.240 --> 00:01:32.440
where you take several models,

31
00:01:32.440 --> 00:01:34.680
ask them the same question, and then literally

32
00:01:34.680 --> 00:01:36.640
sort of vote for the right solutions.

33
00:01:36.640 --> 00:01:39.520
So in some sense, from the engineering perspective,

34
00:01:39.520 --> 00:01:43.480
we're trying to find a way to build a reliable system

35
00:01:43.480 --> 00:01:46.880
from unreliable parts. And that problem we can solve.