WEBVTT

00:25:39.280 --> 00:25:41.800
And to start with, Leonid, one question:

00:25:41.800 --> 00:25:43.800
Is GenAI really new?

00:25:43.800 --> 00:25:45.160
What's different

00:25:45.160 --> 00:25:49.520
with traditional or predictive AI
that we've been talking for a long time?

00:25:49.800 --> 00:25:52.560
Well, I think it's very important not

00:25:52.560 --> 00:25:56.520
to underestimate, but at the same
time not to overhype this technology.

00:25:56.520 --> 00:25:57.560
It's clearly a breakthrough

00:25:57.560 --> 00:26:02.200
from the consumer point of view and from
what it is possible now to do with this.

00:26:02.200 --> 00:26:04.480
But at the same time,
it's important to remember that it's

00:26:04.480 --> 00:26:06.920
actually based on the same statistical principles,

00:26:06.920 --> 00:26:08.440
the same statistical learning theory

00:26:08.440 --> 00:26:10.800
as the rest of AI
and machine learning.

00:26:10.800 --> 00:26:13.200
And in fact, when we talk about generation,

00:26:13.200 --> 00:26:15.880
it is very similar
to the task of prediction.

00:26:15.880 --> 00:26:18.480
Let's say if we think about demand
forecast,

00:26:18.480 --> 00:26:21.160
we're predicting the number that,
for example,

00:26:21.160 --> 00:26:22.400
we’re going to have next month,

00:26:22.400 --> 00:26:24.480
here we’re
predicting the next word.

00:26:24.480 --> 00:26:27.200
So in this sense, conceptually
it is very similar,

00:26:27.200 --> 00:26:29.120
but at the same time to achieve that success,

00:26:29.120 --> 00:26:34.400
if a really important
breakthrough has been made, and,

00:26:34.400 --> 00:26:38.640
first of all, it's a new way of learning,
which is called self-supervised learning.

00:26:38.640 --> 00:26:41.240
It's a way to learn from the data without supervision,

00:26:41.240 --> 00:26:43.200
without labelled examples.

00:26:43.200 --> 00:26:46.880
Second, it's a new architecture,
neutral network architecture.

00:26:46.880 --> 00:26:50.280
It’s called Transformers that has been invented in 2017.

00:26:50.280 --> 00:26:54.960
And then there is clearly an idea of transfer learning,

00:26:54.960 --> 00:26:58.920
the concept that allows you
to train on one data on one task

00:26:58.920 --> 00:27:01.560
and then transfer that
knowledge to other tasks.

00:27:01.960 --> 00:27:03.760
And can you give us examples Leonid,

00:27:03.760 --> 00:27:07.160
on what this is bringing
or changing in industries?

00:27:07.160 --> 00:27:09.680
Well, I can give you example from my experience,

00:27:09.680 --> 00:27:13.120
you know, I work a lot on
predictive maintenance type of problems.

00:27:13.120 --> 00:27:16.640
In those type of problems,
we use numerical prediction to calculate

00:27:16.640 --> 00:27:18.680
the probability of the failure
of the equipment,

00:27:18.680 --> 00:27:21.120
but what we get is really just a number.

00:27:21.120 --> 00:27:23.440
What's important for maintenance screws is

00:27:23.440 --> 00:27:26.760
to understand where this number
came from and what to do about it.

00:27:26.760 --> 00:27:29.400
And so, generative AI
will be able to provide

00:27:29.400 --> 00:27:33.880
the textual explanation of our model,

00:27:33.880 --> 00:27:39.080
plus literally provide us step-by-step
instructions on how to fix that problem.

00:27:39.080 --> 00:27:41.760
Another example could be within banking or

00:27:41.760 --> 00:27:46.160
insurance industry where we try
to predict a risk scores for customers.

00:27:46.160 --> 00:27:49.200
In that case, of course, we're always hunting for more and more data,

00:27:49.200 --> 00:27:50.720
more and more information.

00:27:50.720 --> 00:27:53.480
And 80 % of the world
information is unstructured.

00:27:53.480 --> 00:27:58.120
So this type of algorithm allows us to bring that information towards numerical

00:27:58.120 --> 00:28:01.280
algorithms that actually
predict risk scores.

00:28:01.560 --> 00:28:04.480
IPhone has changed our lives
and the lives of many.

00:28:04.520 --> 00:28:07.320
Would you say that GenAI will be
of the same magnitude and

00:28:07.320 --> 00:28:10.880
that we are going to live another
type of iPhone moment?

00:28:10.880 --> 00:28:15.120
When we're speaking about generative AI, it really depends on

00:28:15.120 --> 00:28:19.960
If you are in the industry, for example,
in car-manufacturing or

00:28:19.960 --> 00:28:23.680
in oil and gas industry, where your main business actually building cars

00:28:23.680 --> 00:28:27.720
or extracting oil generators, generative AI will definitely help you,

00:28:27.720 --> 00:28:32.040
will help you reduce your cost,
will help you improve your productivity,

00:28:32.040 --> 00:28:34.080
but it will not change
overall your process.

00:28:34.080 --> 00:28:36.600
You'll still need to build those cars.

00:28:36.600 --> 00:28:40.280
But at the same time,
if you are in creative industry or

00:28:40.280 --> 00:28:44.240
industry that manages information,
that builds on information,

00:28:44.240 --> 00:28:45.840
generative AI can be
transformative for you.