Understanding Generative AI Fundaments - 15th Mar 24

0:02
Recording.

0:03
Got it.

0:03
OK.

0:03
It's getting recorded.

0:04
Yeah.

0:04
So, yeah.

0:06
So thanks a lot.

0:07
So today we are going to hear from, you know, one of our own people in the India C + C organization.

0:19
They have been you know doing their own investments in the world of AI, conversational AI and
leveraging the ChatGPT wave to bring innovation to the products they are working on.

0:35
Today we are going to look at some of the fundamentals of generative AI.

0:40
What goes in this is think of this as a very basic introduction to generative AI LLMS, the playground.

0:51
How can we generate use generative AI to solve new problems that previously were not possible?

0:58
But more importantly, how to, you know, get started right?

1:02
For those who have not yet crossed the line and played with the playground that's available publicly
or with internal Azure Open AI, how can you cross that line?
1:14
Start, you know, building your first conversational AI agent.

1:19
It could be as simple as answering a few questions that you know you think you know will help your
customer or your internal partners or your Dras.

1:28
This is a fundamentals 101 codes, but we're going to follow this up with, you know, continued
sessions that will take you through the journey of getting into a more advanced, you know, course of
prompt engineering.

1:43
How do you mix search and prompt engineering?

1:47
How do you use plugins?

1:49
How to create more, you know, complex orchestrations when you have to use multiple plugins and
aggregate answers from plugins to answer questions.

2:00
So we have a this is a series of sessions, but today is the focus on the fundamental basics.

2:09
And I'm hoping that you know a lot of you who might probably watching this right now live or who's
going to watch this the the recorded session can actually, you know, cross the line and start your first
generative AI programming.

2:24
And the idea is that, you know, for us to create an AI native organization, all the people here who we
want to have an AI native mindset and participate in this AI wave, you know, getting started with this
generative AI is very, very important, right.

2:42
So this is going to be basics, very simple for everyone to understand and grasp.

2:47
I'm really looking forward to the session.

2:49
Thanks folks.

2:50
Devraj and Beam, please take one.
2:55
Yeah, probably I can get started on that.

2:57
Thanks Vidya for the context.

3:01
Good morning everyone.

3:01
Today happens to be the second session of C plus EAI community.

3:06
Last month we had the kick off and we saw a session on Sydney framework.

3:10
As you might have already seen in the invite like Vidya was pointing out, there are a series of sessions
that has been you know lined up by the V team that we could look forward to both in terms of
learning and presenting.

3:23
Today's topic is Generative VI fundamentals and the mandate that has been given to us is we are not
looking to train people on building different generative VI models.

3:32
Rather we are looking to get into the breadth of offerings that Generative VI brings to the table.

3:39
Enabling professionals like us build solutions and experiences that were previously not possible.

3:45
So irrespective of your previous experience and knowledge on this field, you could look forward to
take something useful and novel away.

3:52
From today's discussion.

3:54
The structure of content looks something like this.

3:57
The 1st 30 minutes probably will send spend on understanding some of the core concepts, the basics
and stuff like that.

4:04
And the next 15 minutes we'll look at some working demos.
4:07
We'll understand how this generative AI situation is shaping up in 2024, what are the advantages,
disadvantages and stuff like that.

4:15
And also if you had a pertinent question in your mind, like if AI is going to take away our jobs, then
probably we'll try and address that through the contents of this discussion.

4:28
And obviously NVIDIA CEO making statements like this doesn't help the situation.

4:34
And someone who is riding the wave of AI and making lots of revenue.

4:38
So let's get started.

4:40
Is my screen coming through?

4:41
Can someone convert?

4:44
Yes, I can see Sasat.

4:45
Yeah.

4:49
OK.

4:49
So is generative AI or for that matter AI, something new?

4:54
I think definitely not.

4:55
Most of us would agree that it is a natural evolution of computer science.

4:59
Humankind started the journey around 1950s with the introduction of first rule based algorithms.

5:06
The first major shift came in towards the 80s where we started seeing the search and optimization
algorithms because of which many search based companies you know came into being and flourish.

5:19
Towards the start of this century.

5:21
We started seeing the early versions of supervised unsupervised models in the field of machine
learning, enabling different scenarios like classification, analysis, recommendation and whatnot.

5:33
Then the, you know pace of the invention, the innovation started really fast.

5:39
Within the next 10 years, we started seeing neural network based deep learning models which helped
in the scenarios of natural language processing, image detection and stuff like that.

5:49
So if you take yourself back by some 4-5 years, you would remember that Microsoft was already
having a lot of cognitive offerings in terms of Microsoft Vision, Cognitive services, Luis, Q&A maker,
even Microsoft tried its hand on putting few chat boards for the public domain like they RU and all
which were not very successful due to various reasons.

6:14
But yeah the point I was to drive, I want to drive home is Microsoft was ready with a lot of these
offerings.

6:20
Then towards the 2017 time period the focus started drifting towards the building different generative
AI models.

6:28
As a result of which we saw the ChatGPT coming into the public domain towards the end of 2022,
which created a lot of buzz and it took the industry by stock.

6:40
So that is the journey that we had seen so far of what what are the aspects that separates the GPT or
the genitive AI models from the previous Tamil models that we had Previously we were able to answer
questions, answer some basic questions, make some object object detection, image detection and
stuff like that.

6:58
But with GPT coming into picture now we are able to generate a lot of creative content.

7:04
Date stories, poems, food, images, videos, whatnot, right?
7:10
You're you would have recently seen the hype around the SURAM model and stuff like that.

7:15
So what exactly is GPTGPT stands for Generative Pre, trained Transformers.

7:23
Generative meaning ability to generate a variety of content based on your input.

7:29
Your input will be in natural language and you know the output you could expect in the form of
natural language images, videos and whatnot pre trained.

7:38
Meaning you do not have to train your models to get started.

7:41
You can start by asking your questions and expect a satisfactory answer because the models are
already trained using a lot of data that is available in the digital media and a transformer meaning self
attention.

7:54
I think we have a slide on attention that the Res will cover, but the high level meaning of self attention
means that the model knows how to focus on relevant stuff and ignoring you know the non relevant
ones.

8:07
For example if you give a very verbose sort sort of an input to your model and ask a very specific
questions, it would give you a very pointed answer to your you know question.

8:19
Ignoring all the non relevant topics that you have provided today what made you know this invention
of GPT or the generative AI large language models possible.

8:31
Basically these are the three major factors the availability of the deep learning architectures through
neural networks.

8:37
Basically there's models are they get trained like a human baby right.

8:43
Baby starts talking about you know it starts talking about Mama Papa, then it forms some words at
the initial stage then it starts forming sentences and eventually it starts making sense.
8:56
So through this neural network based.

8:59
Yeah, sorry to interrupt.

9:00
Sasha, are we on brief history of generative AI slide?

9:03
Because we see the same no slide.

9:08
Yeah.

9:08
Yeah.

9:09
So I'm still on the same slide, sorry.

9:11
OK, go ahead.

9:12
I'll just move to the next one in just one minute.

9:14
Yeah.

9:14
I'm just trying to give up a intro about the GPT.

9:18
Yeah, the point I was trying to come to was like a human BB GPT also gets trained eventually.

9:26
So the GPT 4 that you currently see in offering it is trained using 13 trillion tokens, 1.8 trillion
parameters across 128 to 120 layers of neural network.

9:40
And for all its computing needs, we have our Azure, AWS, all these cloud offerings which are currently
there, which are enabling all these memory and the processing needs of this training of the CPT
models.
9:55
And because of the availability of all the digital data in the Internet, you know we are able to train
these models using a large amount of data.

10:03
That is the reason these are called large language models like you must have.

10:08
Whenever we talk about these models, we talk about big data, large language models.

10:11
So it it requires a lot of data to get trained and because of which the data that is available on the
Internet we are using that to train these models.

10:21
Is the evolution ended?

10:24
Obviously not.

10:25
The next will be artificial general intelligence where the machine will have a philosophical purpose of
existence.

10:32
They will know what to do next.

10:34
So right now we are asking the model or the system to do something for us.

10:38
Eventually they will know what to do next based on the purpose of existence.

10:43
We have taken this small meme from this movie.

10:47
Dune Dune 2 is recently released in theatre.

10:49
So if you remember this character who is Amentat who does a lot of, you know, heavy processing
workloads, within seconds they are identified.

10:59
They live a normal human life.
11:00
They are identified by a tattoo on their lips.

11:03
And yeah, that's just, you know, glimpse of how the future might look like in the form of artificial
general intelligence.

11:11
Moving on to the next slide, what is a foundational model, the core of any a generative VI is a a
foundational model.

11:21
GPT, you know being trained with a huge amount of data and being able to adapt to a lot of
applications fits the bill.

11:29
It is able to cater to a lot of a variety of you know, scenarios like answering questions, image
generation, summarization, whatnot.

11:38
With that, I'll probably take a pause and hand it over to Devraj to cover the, you know, concepts and
stuff like that.

11:51
Devraj, thanks Saswat for briefing the history of AIML at GBT.

11:58
Let me take over the screen and meantime if somebody has any question please feel free to you know
ask.

12:14
So for now 30 minute what we're going to cover is the basic of you know fundamentals of sanity, VI
and it really helps to understand the fundamentals of anything, right.

12:27
Like for example today if you are working on any product for example you are working on a sequel
internally.

12:34
How the data is you know managed saved into different table data pages on disk.

12:39
That really does not bother when you are just writing a query or retrieving some data.

12:44
But let's say if you have to design A complex system then definitely you have to worry about the
indexes.

12:50
What kind of indexes you need to build in database right?

12:54
Cluster, non clustered.

12:55
So internals definitely helps when you start building products and like system or complex system that
is where the aim of today's process and primarily you want to keep fundamentals, the insights of the
AI or transformer model, what we call using today, right.

13:14
So here are the topics what we're gonna cover.

13:16
So I will just brief about the applications of AML which we are using in day in day, day, day in life right
just to connect it, what we feel where AML is getting used.

13:28
Then some brief about CNN and transformer and why the need of generative AI evolved.

13:35
And then these are the different terms which you would have heard in previous you know various
sessions embedding attention transformation multi head and how internally all these things work and
why they are making your transformer model work or excel.

13:52
What happened suddenly in like 20, 2324 that all are talking about AI, what was, what is new in this
right that that's all we would try to discuss and go through.

14:07
So here are the applications all AML related to use stay in and day out.

14:13
And sometimes we don't realize OK these are powered by your AIML social media filters.

14:19
We we keep using that right in different platform like Instagram, Snapchat, all those are again
powered by your AIML stuff recommendation system.

14:29
You buy anything on Netflix for movies.

14:31
Spotify, again the AML language translation app is again powered by same AML, the copilot and
Virtual assistant which we are building today, you know on different applications.

14:44
Again the power of AIML coming there and especially the generative AI is boosting that.

14:51
And the photo editor app also we we use in our mobile, right.

14:55
So all those are again powered by your AIML stuff.

15:02
OK.

15:02
So in this slide what I'm trying to cover is what is new which made transformer or generative AI Excel
today.

15:12
Previous to that as Saswat covered in the brief history, we had all those stuff in the market from you
know 1980s, nineties and 2000 also to 2017.

15:24
Everything was there in place but there there are some shortcomings were there.

15:29
Like for example if you go to neural network there are right side you see a feed forward neural
network where everything is get reminded by the network all the knowledge whatever requiring your
neural network is gets preserved in the network itself.

15:46
Different node will hold those values or knowledge I would say.

15:50
Whereas recurral recurral network was next step to that where it is able to you know feed backward
and forward which enabled us a time series kind of support.

16:02
And then came CNN right where you can your face recognition kind of system number recognition
from distorted images and all it.

16:12
It has this layered kind of architecture where it can knowledge like remember and then classify and
predict as well.

16:21
So that if you look at the left side again the drawback what they're like I would say primary drawback
if I say in RNN and CNN is the lack of parallelization.

16:31
They you cannot parallel, you have to train your models sequentially.

16:36
If I give you just a glimpse of that like for example today SAD GBT is trained over around 45 plus
petabyte of data.

16:46
With the same amount of data we would have need to train a RNN.

16:49
It it will take approximately I think a couple of years not even months so, so think about that if you
need a model to train years then will it be usable because every day new information is coming into
market right?

17:03
That is where even though if you train the same RNN with like same data which today your chat GBD
is doing, it will give you the same kind of answers but again the training cost and the time lag which is
not making it very useful right with the latest information at all.

17:23
That is where these neural network were lacking.

17:28
They were not supporting the parallelization to do the training.

17:33
Same thing with the CNN also it was lacking in terms of the same feature, not able to parallelize and it
can just remember the information in network.

17:44
There is no backward or forward kind of propagation.

17:47
Backward especially propagation of knowledge possible.

17:51
Whereas in transformer it the IT enabled you do the parallel training.

17:56
Training the parallel capability actually made it excel.

18:00
So in coming slides you will understand why it supports the parallelism which is where the
fundamental will help you better understand.

18:08
That was the primary reason making this transformer you know excel in in generative VA.

18:16
You can go in other points in detail from the slide later on.

18:20
But yeah this is it.

18:22
So what is large language, large language model I think has already set the context in that.

18:28
So I will just glimpse through this slide.

18:34
Here is the architecture of your transformer.

18:37
You would have seen in different sessions.

18:38
Right, left side your encoder, right side your decoder and you you pass in the input to your
embedding then output and these are multiple boxes in place.

18:51
What they are doing and how they are actually enabling your transformer to work is we will cover in
detail.

19:01
So what we do in generative AI, Generative AI we generate the output.

19:06
You ask a question actually it is answering your question.

19:09
It is generating the information for you, right?

19:12
For example, if you are asking a simple question, the sky is the.

19:16
If you ask that question, the next word, what it has to fill your LLM.
19:21
The last language model will try to find it out and put you the best word as the answer to that prompt
to complete that sentence, right.

19:30
So it will be putting it.

19:31
The sky is blue.

19:33
Now why blue?

19:34
It it could have put you know red or orange is the sky is orange, sky is red.

19:40
But why?

19:41
Why it made it like blue is the best fit to it.

19:45
That is where LM uses internal of internals of that mapping and all that will come to know into next
slides.

19:53
But just to give you an test of that what LM is really doing.

19:58
Sorry, somebody has any question?

20:01
Hey there Raj, this is Rohit.

20:04
I'm just trying to understand like.

20:07
So this was the sky is blue kind of an example already existed when we had transformer models or
before the different kind of model, right.

20:16
So what has LLM done differently to this versus what it was to be do earlier, correct.
20:22
OK, that is where I think you're asking difference between RNN.

20:25
RNN also could answer you these same thing.

20:28
You are right, right.

20:29
But the problem is this is just a simple sentence of three word four word.

20:34
When you train your model, let's say with entire data world, let's say Wikipedia, then the correlation
and context.

20:43
The RNN is not capable of handling that.

20:45
You can maybe it can remember the context maybe till 1520 words.

20:52
Beyond that it loses the context because the training in RNN is sequential.

20:56
It tries to map your fifth word with the first word and that is a sequential training.

21:03
That is where the the next concept which is coming, we are discussing about the attention which
attention actually overcomes their limitations of remembering sequentially.

21:14
So it is good like Arun and other can answer simple sentence.

21:17
But in long sentences if you have to train let's say 10 pages of document and then you answer a
question, those model fail actually.

21:26
Got it.

21:27
Thank you.
21:28
OK.

21:29
Just to add a little bit more context, Right, Rohit, to your question, sure.

21:34
So the auto encoder models is just a particular topology of a neural network, right, Right.

21:42
So basically at a high level, neural networks are nothing but multiple layers of neurons that learn a
certain, yeah, particular weight and bias.

21:54
Right Now when we use, you know, we talked about RNNS, right?

22:00
RNNS are recurrent neural Nets which have some, you know, additional capabilities.

22:07
And then we had LSTMS.

22:09
LSTMS added memory, right.

22:12
It's not just learning on the family, but it's keeping some memory so that it can use that memory to
refine its learning.

22:21
And then after LSTMS, the same LSTM topology, network topology, there was some nuances called
attention based models, right?

22:30
So these are nothing but improvements, very incremental improvements, improvements on the the
topology of the network and how it learns.

22:41
Now LLMS are nothing, it's the same concept, just that they've been trained on humongous amount
of data with explosive number of parameters, right.

22:54
So if you remember a few, 5-6 years back in, we had Google launched something called Bert, right?

23:03
And then we had a series of innovations on top of it.
23:06
So they were all also called.

23:08
They can also be termed as LLMS only just that we are in the latest generation of LLM where the
number of parameters has gone into.

23:16
I don't know what is the latest in chat if I keep forgetting the number into thousand, hundreds of
billions rate of parameters.

23:23
Previously it was not even possible to train those kinds of models just because of the raw, you know,
lack of power in terms of GPU's in terms of the distributed forms that were not available.

23:35
So this the network topologies have incrementally improved, but what has drastically improved is our
training the farm right.

23:45
The kind of machine pool we have now and we are throwing at this problem is the real reason why
this has become extremely refined and almost human like in its language kits.

23:56
So Vidya this is useful just rephrasing in a little bit simpler term if you can, right?

24:02
So we are basically saying that the the encoder, decoder is more or less same the neural network
topology of how to connect the nodes and find the vectors and everything is still the same, it's just
the input has exploded in multiple folds that helps us to make that encoder decoder on the neural
network to operate more efficiently than it used to do earlier.

24:26
Yes, in simple terms that is it's just the amount of data we are giving an amount of times we are
training the Epox and the number of layers we have added has just brought out the human like
behavior.

24:37
Oh thank you.

24:38
So thanks Vidya.

24:40
And in simple terms just to add, if I just let you know that if in case of RNN when you are training and
you have let's say different words, they can correlate to each other in sequential order.
24:53
So it has to remember what is the worst word previous to it, previous to it and if it has to correlate
with the last word to first word.

25:01
Let's say you said how many fruits are there in the market.

25:05
So the first word fruits in market they will correlate with each other sequentially like it has to go
through this word, this word, this word that way.

25:13
Whereas in case of transformer the the concept which we are going to discuss is attention, attention
all what you need what it they change the architecture in such a way that let's say you have a word in
the sentence instead of having a sequential dependency.

25:28
Now the mathematics behind that the calculate is they create a matrix by by by by calculating the
distance of that particular central.

25:40
I would say matrix one in the the distance between your number with what you are going to calculate
the attention.

25:48
Attention is equal to all the words.

25:51
There is no loss in terms of the length of the sequence.

25:55
Whereas in case of RN the loss will be there if there are who's in large sentences.

26:00
Whereas the way they the change the architecture that they calculate similarity between different
sentence by taking a central calculation which each word will have correlated to that only.

26:11
So there is no loss in terms of the length of the word.

26:14
So just to let me go back to slide now.

26:23
Yeah, so how LLM here you see that these words are getting translated into some math.
26:31
Mathematics, right, Because your machine cannot understand the alphabet in numbers, sorry,
alphabet or characteristics or English, right.

26:40
It has to get translated into numbers.

26:42
Now how to translate those English or any other language characters into number that is all we call it,
that is.

26:52
Sorry embedding.

26:54
So embedding is nothing but it is translate translation of your English.

26:59
I would if you consider the language English, then translation of that English into numbers is all
embedding because your computer can understand numbers right?

27:08
It cannot understand your phrases or English words.

27:12
So the embedding is just converting of those into numbers.

27:17
So you would have heard these sentence this word multiple times, right?

27:20
This is all embedding is in simple term that's conversion.

27:27
Now this is just one example of that.

27:30
Let's say you have different items in place with you.

27:34
There are two axis where it has coordinate X&Y and Now if you look at these objects and the right
side there is a table.

27:44
Each object has a coordinate.
27:46
There's just the X&Y coordinate, let's say car, car horizontally 6 distance from zero and Y axis is at 0
right?

27:57
So that is how X and YS.

27:59
So that that's how the other words if you see all has coordinate X&Y.

28:06
Now what you have done here is you have actually converted your your objects into matrix which is
embedding.

28:15
So converting of your real life information in different language into numbers is all embedding is.

28:23
So let's see if I ask you a question where the apple should go.

28:27
Like Apple, you have to embed the apple, let's say with.

28:30
Given this as a context, where should apple go?

28:35
Should it go to like this group, this group or this group or this group?

28:40
So coordinates finding that coordinate is all embedding.

28:46
If somebody want to give a try with yeah, hand up.

28:50
Oh sorry, yeah, go ahead Vidya, sorry, I could not see that things.

28:58
Yeah.

28:58
Vidya.
29:02
Yeah.

29:02
I think it's from previous, so.

29:03
Oh, OK.

29:04
OK.

29:04
So, OK, thank you.

29:06
So thanks I for letting know.

29:13
So apple will go here, It's a fruit, right.

29:17
So coordinator 55.

29:24
So what?

29:25
Embedding.

29:26
So the embedding.

29:27
Right now what we have seen is in X&Y matrix.

29:32
So here are the words and the numbers are the matrix the embedding of those words.

29:36
Now these are very simple embedding which helps us to understand OK you know what is really
embedding is.

29:44
But if you go to real life data or example which model has to learn, it will have lot of embeddings for
you know for translation of your real life features into that into machine which can understand.

30:00
Like for example in previous one your banana, banana what is the colour of the banana?

30:06
That will also has to be translated into some some numbers which a machine can understand right?

30:13
So what it will it will do internally?

30:15
Maybe it's like it will have one more column having some numbers to that which will denote the color
of the different different fruit, those fruits, right.

30:24
So the number of parameters will increase that is where you keep hearing.

30:29
OK, some models are huge in parameters like I think 250 billion plus parameters.

30:35
So this is because we want to cover different aspects of same maybe those entities what we are trying
to learn or embed which machine can understand because we don't want to lose any features in that
translation.

30:51
So we keep adding you know more columns or rows to the metrics.

31:00
OK.

31:02
There was there is a question what does 4096 mean in the previous slide?

31:07
OK yeah.

31:08
Oh yeah, this is just the length of the metrics, what really we are using most of the cases in today's
high end LLMS, just the length of the metrics basically that or you can simply say the features of your
your products or your fruits.

31:26
So how many features you want to record into in this embedding which a machine can understand?
31:32
So is this specific to this example or are are you say stating something more generic?

31:38
No, it is specific to this but these numbers.

31:41
This is an average what we are considering today.

31:43
But it can be changing.

31:44
It can have lesser mode.

31:46
As I said, these are just the feature which you are which you want your machine to understand, right?

31:51
Let's say you don't want to you are not bothered about the color of the fruits, so you will drop one
feature, so you will have one less column in that case maybe.

32:01
OK, sounds good.

32:02
Thank you.

32:02
Add it to the parameters David, each column can I say it as mapped to the parameter.

32:11
Yes, yes that that that it's a that is a good definition.

32:14
I think to remember this what what these features in the metrics it is little more in complex but they
are in basic terminologies.

32:22
We can that is the right thing to understand.

32:24
I think you are right on the number of parameters of of a product or item You can have more and less.

32:37
So again another example you see different here right items and then they are also arranged in the
X&Y matrices.

32:49
All the fruits are going to right top and all the techie products are going left to bottom right.

32:56
Now where should apple go?

32:58
Apple is a fruit as well as your technology brand as well, right?

33:04
That is where the problem starts.

33:07
If there are simple, if you take the simple embedding one or two columns, they may give you clear
answer.

33:17
But if you have to remember more, if you have to answer more complex questions in case of
especially ambiguous products or items, then the problem is there.

33:30
So what is the solutions there because it can go both the side right.

33:39
So we'll come back to this question in next 2-3 slides.

33:43
Before that we talked about embedding right.

33:46
So is the is 1 embedding good for your entire machine learning?

33:53
The answer is no right?

33:54
Ideally we want to have more embedding and the and the type of embedding what is good or bad.

34:01
So for example here are there are three examples of embedding where you have your phone and fruit
and apple in between another same where you have apple then then your items are here fruits and
phone and again this are here right?
34:16
So what is good and bad embedding?

34:20
Can anybody make a guess here which one will be good or bad?

34:30
Because you are the first one.

34:32
Yes right?

34:35
So first one will be the good one because it is giving you clear separation of if you move that item to
either of the side, they will be closer to a particular context.

34:46
So it will help you, you know remove the ambiguity.

34:52
Whereas if you take other two, like if you move a bit of your coordinates, maybe we we saw like X&Y,
if you move the coordinate by a little bit it's still the the context is ambiguous so you may still get the
wrong answer.

35:06
Whereas in case of first it is good.

35:08
So likewise this is also bad and this is a good one.

35:12
Now embedding is very crucial, we have seen it here right?

35:17
Which really drives the answer how any LM can give you.

35:21
So if it is a good embedding, it will give you good answer like if you're asking questions to chat GBD,
if that training is done with a good embedding it will give you very good answer but if it's if it is not
then the answer will not be that great.

35:35
So entire power of your LM is getting driven by embedding.

35:40
This is we we we like, we hear very less like we talk about all other terminologies right.
35:46
But this is the core and the powerhouse I would say of any LLM.

35:50
How good you are embedding how you'll get more good.

35:53
The answers are so.

35:56
So one embedding will not help you get that clear answer because if your parameter increases like
here we have just taken 2 dimensional.

36:05
But in real life there are multidimensional multiple hyperparameters.

36:10
Right solution is to build more embedding then so that we can clarity we can we can get the clarity in
terms of how the separation can be done better.

36:21
Now creating embedding is not that easy.

36:25
That is where the computation power of any LLM goes, because you have to translate one language
denaling least into numbers.

36:35
So how to get new embedding?

36:36
That's another terminology.

36:37
You would have very often heard linear transformation because every LLM does that because one
embedding cannot give you clear cut answer.

36:47
And if you go and create new embeddings, the power required for that is the huge like you need
more GPUs in terms of the generating those embedding then getting train entire network.

37:02
So easy way out that what we do is linear transformation, where let's say you have a image, you just
rotate it by a bit of angle.
37:12
You stress it in X or Y coordinates.

37:17
Even though it is the same image, it means same embedding.

37:21
But you change coordinates to make a different view of the same image and you get new embedding
out of the same embedding.

37:30
So you don't have to do a recomputation of, you know, transfer, translating your English again to
numbers.

37:37
So there we saved the some power in terms of computation.

37:40
That is called linear transformation.

37:44
This is just a.

37:46
Yeah, just a time check.

37:47
Sorry.

37:47
Yeah, 10:40.

37:51
OK.

37:52
So I think another 5 minutes.

37:54
Yeah.

37:56
So new embedding from existing one, right.
37:59
So here is another example of where we have creating embed.

38:04
Again we just rotate the the coordinates and you get new embedding and there are good embed
embedding again based on the length right?

38:15
So what we do is we give a score to those.

38:19
Now we go one step further not only saying good and bad but we also giving a score to your
embeddings.

38:26
That's the first one.

38:27
We give a score one 2nd 1.01 because it is little ambiguous and not clear enough.

38:34
If you try to get answer of the answer out of this, the 4th one is it is still good.

38:40
So score it's the best I think out of the all three of them.

38:43
So score is 4.

38:47
Now in your transformer model you would have seen this again this scale dot product, attention and
concatenate and linear briefly.

38:57
I will test upon this, but internally what all these these boxes are doing are creating the embedding
from by linear transformation where you see the linear one creating multiple transformation from the
same.

39:11
Wherever you see linear bang on.

39:14
It's a linear transformation.

39:15
That's it.
39:15
Nothing more.

39:17
Scale dot product is just that you are giving scores to those transformation.

39:22
You are scoring those transformation.

39:25
That's it.

39:27
Nothing fancy.

39:28
Science.

39:28
Science is happening here.

39:29
You just remember you are this scoring your transformation which is good and bad.

39:33
That's all it is scaled and the concat in it.

39:37
When you combine these transformations right into one because you have to maybe penalize or
remove the transformation which are less helpful to you and which are more helpful give emphatize
or give priority to those.

39:53
So you concatenate those transformations that's all the concat, concat this box is doing.

40:00
So you know all these terminologies now by just looking at this image by embedding.

40:08
Now attention we talked about earlier, right, why the transformer got success and I just also sold one
image where RNN giving a sequential.

40:21
Dependency of to remember a context what was said in the beginning, what is the middle and the
end.
40:28
Whereas transformer has new mechanism attention which is I think in 2017 the document was
released by Google and based on that and this transformer model was built.

40:40
What is attention?

40:42
And in in your transformer model these are some math behind they do is like took.

40:47
They actually calculate attention.

40:50
Yeah with the input, whatever you are passing these input, the transformed metrics of your input
itself, the linear transformation which you talked about.

40:59
Right.

41:00
So now what is attention in your sentence?

41:05
If you look at like, please buy an apple and an orange and versus apple unveil the new phone.

41:11
Right now your apple is going to two different contexts.

41:16
One is orange is a apple is a fruit where is apple is a brand.

41:21
So who is deciding that?

41:23
What is helping that?

41:24
That is the context which is helping that.

41:26
That is where the attention given by the orange to apple and in this case the phone is giving attention
to the apple.
41:34
So if you're asking a question with OK, what is Apple phone price.

41:39
So then the phone will decide, OK Apple is is a brand technology brand for device not a fruit.

41:47
So this is all the attention is given by other words to whatever the primary word I would say you want
to find it out.

41:56
Now you see you have seen here that tension by giving last word to middle word and then there is
just one-on-one length between these two words.

42:06
Whereas if you go to again RN and CNN they they have, they cannot give like this attention to one of
the last word, cannot give attention to the middle word.

42:14
They will have to remember the last word, will give attention to previous, previous, previous and by
the time they get reached to the end, we lose the context.

42:22
In case of the previous neural network, that is where they are not, they did not succeed in terms of
this.

42:30
Now just a graphical representation of the attention again.

42:34
Now you have the coordinate of orange and phone here right now the attention given by these two
words can take your apple either to towards the phone or either towards the orange.

42:50
So your coordinates of the apple launches because of the attention attention given by the other
words.

42:57
So new coordinate.

42:58
You can decide whether if it is a phone is coming, the new coordinate will be accordingly shifting
towards the bottom side and the the attention is from orange.

43:07
Then the coordinates will change towards an upper side of the fruit just to you would have seen this.
43:20
I think Saswat may cover detail here but just to give a glimpse like Azure Open eye when you create
resource here and you can open go to studio chat and this chat prompt, you can just go ahead and
ask simple questions and you can add your data as well here.

43:40
OK, whatever data you want your model to correlate and use.

43:44
But if you just ask a simple question here, it will help you answer and you can see example of let's say
let me simply ask a question what is the cost of Apple phone?

43:56
Apple phone?

43:57
Now if you look at this because the attention given by the phone, it knows that Apple is a brand name
for technology.

44:06
Whereas if I ask the second question attention changes because of the the word it takes that as a fruit.

44:18
So it is little ambiguous.

44:20
Ambiguous right now.

44:21
But you have to clarify.

44:23
It will ask the clarifying question in case if it is ambiguous in nature because then as I said the metrics
which is try to calculate it has to get get to a clarity.

44:35
If the clarity is not there, the way the chat boards are built, they ask you more questions.

44:43
Now what is pulling these objects to either side?

44:47
Right.

44:48
So that is called.
44:49
That is all called as context pull in similarity.

44:57
If one word is sitting there then all the other words actually pull that first towards it based on the
based on the similarity which it has with other words.

45:12
Now how do we calculate the similarity?

45:15
Right similarity calculation is what we have seen which we call this dot product or cosine similarity.

45:23
You would have heard these two words as well.

45:26
In case of dot product, let's say I want to see whether any of the item is falling towards phone or
towards fruit.

45:33
That you can simply you have calculated right their coordinates and then if you multiply these
coordinates and then there will be a number coming out of that.

45:45
If that number will decide OK, what is more similar towards for example orange and phone, they are
different apart.

45:54
If you multiply their coordinates they will sum up to 0 because they are very distinguished.

45:59
That is where machines should also understand and learn.

46:01
Whereas the other words in between they have some similarity.

46:05
If you multiply their coordinates they will have some number.

46:10
Likewise cosine similarity is just when instead of using the coordinates we use the angle angle
between like a distance.

46:17
For example orange and phone, we have angle of 90 so you take a Cos 90 so Cos 90 is 0.
46:26
So there is no similarity between these two items whereas you take the angle between any other like
these two products are cherry and orange.

46:33
The angle would be something like this I think 14 or so.

46:37
So it is very similar because the angle is less and the cost of that is coming point 9.97 which is like very
similar.

46:44
That is how machine internally calculates.

46:47
You know how a product is similar to or a word.

46:52
I would say a word is similar to the word or not.

46:57
Now earlier we just talked about the simple example where we had to coordinate but in real life just
this is just giving idea there there will be so many coordinates depending on how many parameters
you are adding.

47:08
So number of parameter increases your your you know all what calculation will also increase in terms
of the embedding right?

47:18
You have to add more column and rows to understand that and code that in terms in terms of the
machine.

47:26
So all these coordinates and that is where we call hyperparameter like number of hyperparameter
increases because then it has to learn and understand and keep knowledge of those different
coordinates into its matrix.

47:44
To end with that, I think the soft function I will just end the fundamental part before I hand over to
Saswat.

47:50
So soft function is the last function which we have seen in transformer model here.
48:00
So starting with input embedding, linear transformation and is the softmax Max.

48:07
Now when you give a particular word for your transformer model to answer, it may answer different
words.

48:17
It may answer maybe when you say what is how is sky is as I said earlier it can give you answer of blue,
red, orange, red.

48:28
Now what is what is the best answer to that?

48:31
Because your model will generate the generate the generate the scores, what is?

48:37
What is the good embedding?

48:39
What is bad.

48:40
But to translate that knowledge into the probabilities, all the soft softmax function does.

48:48
The purpose of softmax is to adjust the output of a neural network that they sum up to 1 so that that
that is how you know like which one is the best one.

48:57
So if that is let's for for example this here .57 and .43.

49:04
So if this is a higher number they have one scale.

49:06
Now if I say OK from zero to 1 which is more closer to me, maybe if I take higher the number scale
then this is better answer to that.

49:16
So softmax at the end of transformer just translates all the output into probability between zero to 1,
so that you can pick the one best answer out of that.

49:28
So here are the fundamentals, I think we had less time, so I just skimmed through a little faster in
terms of different terminologies, but feel free to reach out and ask question, we can have more detail
you know discussion later on coming sessions we'll we'll we'll cover any anything specific you want to
you have added as part of the fundamental.

49:50
So I think now to that I will hand over to Saswat, we'll be covering the building blocks of a solution
based on Zener DVI.

49:59
Hey thanks Devraj.

50:02
I think there's few questions on the chat if you want to take a look on that.

50:07
Yeah, if someone can confirm again if my scheme is coming through, yes, I can.

50:17
I can see Sasotiya.

50:18
OK oh, so we have another 7 minutes.

50:21
We'll try and cover the rest of the things.

50:23
Yeah.

50:24
A typical, you know, solution that is based off of generative AI would look something like this.

50:31
So let me just not this event.

50:39
Yeah, at the very the forefront of that you have an end user interface, right similar to your ChatGPT
where you could, you know, type in your prompt and you could get your response and stuff like that.

50:49
It could be your chat bot, it could be your Pi powered application, whatever it is.

50:55
Below that there is an, you know, So there are a few animations here.
51:00
Yeah.

51:01
So the, as we said Copilot interface, user interface is at the very forefront and below that there is an
orchestration layer.

51:11
So in the last session I think we had a discussion on the Sydney framework that is one of the
orchestration platforms within Microsoft.

51:17
We are encouraged to use semantic kernel and outside you know and there is a famous open source
framework for Lang chain.

51:25
Recently Lama Index is also making a lot of you know Wave.

51:29
So what exactly is orchestration layer?

51:32
Basically it you know.

51:34
It knows how to interact with your rest of your ecosystem.

51:38
It knows how to convert your natural language text to your embeddings and stuff like that.

51:44
It knows how to interact with the storages that you have in place and stuff like that.

51:49
It is basically a set of agents, skills and plugins that are executed either sequentially or parallely in
order to give you the outcome of your prompt that you have provided.

52:00
We'll have a session detail on this in the afternoon so look forward to that.

52:07
And how the orchestration layer interacts with the multimodal large language models is basically
using the prompts and conflicts.

52:16
The ones that the was just displaying the prompts as nothing but the natural language text that goes
and input to the to the model.

52:26
And you know again it can it could respond in the form of a natural language, it could respond in the
form of you know image or whatever it is.

52:35
However not always you know we will use the text as an input to our model right?

52:44
Sometimes we could use images and audio and stuff like that.

52:47
So because of that we have a multi modal system that basically supports all these features.

52:53
So if I take you to the playground of Azure, you know, yeah anyone has a question, OK, something
yeah.

53:06
So along with the prompt that you will type in here, you could also set the set of parameters which are
nothing but your configurations.

53:13
You will basically tell the model on how much creative or con size it has to be.

53:18
So basically a combination of these two helps in effective communication with the model itself.

53:23
That's nothing but a combination of second.

53:37
Yeah, the prompts and conflicts.

53:40
Below that, any AI powered system is powered by multimodal large language models.

53:45
Why exactly it is called multimodal?

53:48
Again, that is because it is a it knows how to combine different types of information, be text, images,
videos, audios, sensory data and providing human like response, right.
53:59
And at the beginning we talked about foundational models.

54:04
So right now the latest ones that are offered are offered from Open AI chat, Jiggity for for you know,
textual data for Dally tool for the image generation, codex for code generation and stuff like that.

54:18
On top of that, you could basically aid the, you know, foundational model with all your cognitive
offerings.

54:26
So Azure supports Vision APIs, cognitive search APIs and there are also a trend going towards small
language models from large language models.

54:37
So anything that is industry specific or company specific that a particular small model that could be
trained on and found.

54:46
Tuning of the foundational models can also be done.

54:49
Just to give you an example, this is you know, if you might have already seen there is something called
Chat PDF where you could provide your input in the form of a document.

55:00
So in this particular case I have provided the annual report of Microsoft as an additional input to this
to this web app.

55:10
It is already having an interface with the Open AI and I'm asking very pertinent questions like extract
the key takeaways from this document right.

55:19
As you know that ChatGPT is trained in the data that was available till 2021 and I'm providing
additional data in the form of an you know PDF document and I'm asking question on top of that.

55:30
It is able to give me answers, you know concise answers.

55:34
What are the from this?

55:35
So that is where this you know providing the additional context and all these things comes into
picture to give you an example of SLM or the small language model.

55:45
Let me again due to one of the demos that we had created, so small language models are nothing but
you know these are, you know, normally smaller by at least 5 to 10 times than the NLMS.

56:01
It takes less energy and less effort to train and you know the computational, the memory needs in
order to train them up also smaller.

56:11
But again, it has its own limitations.

56:14
In this particular case, if you see I'm using Moon Dream which is a small language model and I had
asked a question about this particular person who is this right?

56:24
So here if you see my prompt says that you use this image and tell me who is this.

56:31
For any large language model, it would be very easy to answer this.

56:34
But because it is trained with a lesser amount of data, it could it.

56:38
It didn't tell me that it is MK Gandhi directly rather it what it is telling me it is a mere man wearing
glasses and serious expression and model.

56:46
Right.

56:47
So the point I was to drive home with this example is SLMS are trained with lesser amount of data.

56:54
The IT has to be used judicially, judiciously and if you have any specific data that are put into your
application, your organization and stuff like that for that it would be used.

57:04
Then coming back to you know the Vision API again for that also we have an example, click close this.

57:22
So here what I'm saying is I'm giving an example of you know an image.
57:28
So this is my image basically.

57:29
Let me expand that a little bit.

57:32
I have given it a invoice which has the, you know, GST calculations and I'm also giving it some sort of a
context saying that this is how the GST is calculated.

57:42
And now what I'm telling the model is basically you know, use the description that is provided within
the green rectangle to verify the value that is in the red rectangle is correct or not.

57:55
So green rectangle has my you know calculations and red rectangle has the actual values.

58:00
So it does the calculation, it understands the image and it gives me advancing that.

58:06
We cannot ascertain the accuracy in CGST and GST values.

58:12
And if you see here we have given an example.

58:16
In this particular example, we are using the Vision API to aid the OCR process of this model.

58:26
Open AI provides you know comes with its own vision capabilities.

58:30
However, you could aid and provide a better input using the Microsoft Azure Vision endpoints and
help the process give you a better.

58:40
You could always take out this part.

58:43
I could always comment and ask the same question.

58:45
It could still give me some answer but it may not be accurate.
58:47
Sometimes it says that if it is not able to read the text and stuff like that.

58:50
So that is where your Vision API comes into picture.

58:53
Again coming back to fine tuning, I'll have to take you to our Azure AI Studio preview.

59:00
That is something called fine tuning.

59:02
Right now there are I think limitations on how many models that we can find you so Lamar to be 270
V Lamar, whatever.

59:09
Like all these models are provided for fine tuning purposes if you go to the Lamar models.

59:17
So there is something called explore.

59:19
Right now Azure supports all these sort of LLM and SLM models.

59:24
I can search for my Llama models here which is basically developed by Facebook and open source and
you see this.

59:31
This is a small language model that is available.

59:33
I can go here and it should have the capability for me to fine tune it.

59:39
So you see this tab.

59:40
I can use this tab to fine tune this model.

59:43
So any specific data that you have for your organization, you could use that to fine tune and stuff like
that.
59:47
Now coming back, there is something called Control LLM.

59:56
In the interest of time probably we'll skip this part.

59:58
But this basically helps us, you know, let's say you are giving a very vague sort of an input and you
know the model needs more data in order to analyze and provide your response.

1:00:08
So in that particular case, Control LMS help.

1:00:14
Below this entire thing, we have a you know.

1:00:18
Cloud computing infrastructure for Open AI, we have Azure even I think Microsoft has done some
investment with Mixture recently and that has run into some sort of legal issues.

1:00:27
But yeah, mixture will be trained on top of Microsoft Azure.

1:00:30
If you heard the news about Anthropic AI Cloud Three, you know models showing that.

1:00:42
So this is the company behind Cloud Three and what they are claiming is their efficacies are better
than you know even the GPD 4 and stuff like that.

1:00:52
They are providing a real time calculations.

1:00:54
This particular you know company or the model is based off of Azure.

1:01:01
Sorry, Amazon Web Service, cloud platforms or even Google has made some major investments in this
now.

1:01:17
Yeah, your orchestration platform can talk to your ecosystem.

1:01:23
So in this particular case, probably we will have our Azure ecosystem which is a combination of
keywords, databases, whatnot, right.
1:01:31
So this, the orchestration knows how to use this in the form of plug insurance and stuff like that.

1:01:36
We'll again, we'll cover all this in detail in the afternoon sessions.

1:01:41
And then I wanted to cover was the retrieval augmented generation basically so the vectors that
Devra's had shown basically we could ingest our data.

1:01:51
So the example that I showed here, basically my entire PDF document got converted into vectors, it
got stored into a Azure Cognitive Search API and finally whenever I'm you know providing a query this
my query gets augmented with the vectors that it retrieves from this cognitive search.

1:02:09
Then it sends it to the open AI to give me a natural language text response.

1:02:14
Also if you see here it ports different parts of this article.

1:02:18
So if I keep on this page number the you know the data that it is giving me it backs or grounds that
data based on some sort of you know pages that I had provided.

1:02:30
So that is about that.

1:02:31
I think we are already about 5-6 minutes overtime.

1:02:36
I just wanted to check do we have another 5-10 minutes to cover the last slide or shall we call it a day.

1:02:42
We still have audience maybe.

1:02:48
OK, few minutes I think.

1:02:49
Yeah, sure, sure.

1:02:56
Slide, Yeah, we tried to do a SWOT analysis of generative AI in this particular slide.
1:03:05
So we had that question right.

1:03:07
Is EI going to take away jobs?

1:03:09
Probably the slides would help us answer that.

1:03:13
The strength that generative I brings to the table is it will automate all sorts of mundane tasks that
you know people have been doing.

1:03:22
It is helping and doing quick swift data analysis, real time predictions, you know, So anything like if
you want to do a preventive measure for your security reaches for healthcare systems and stuff like
that, it can help.

1:03:36
It can provide an efficient incident response system.

1:03:38
So any customer can provide a natural language, you know problem statement and we will be able to
understand the context and provide a quick response to that, accurately identify the patterns and
anomalies.

1:03:49
So anything the fraud even you know if you want to scale up or scale down your system based on the
load and stuff like that, we could do that based on the people that is provided by generative AI and
extract actionable intelligence, right.

1:04:02
You must have seen the copilot experience in teams wherein for a meeting just given a transcript is
able to generate some sort of a tasks for everyone in the core.

1:04:12
Now going to the weakness side of it, the major problem with generative AI are the hallucinations.

1:04:22
So basically generative AI, the last language models, even the small language models does not know
how to say no.

1:04:28
So the example that I gave here, I asked a very basic question.
1:04:32
Who is the person?

1:04:33
And any you know layman person could tell me or any you know minutely moderately trained model
should be able to tell me that it is MK Gandhi.

1:04:42
But it gave me a very vague description.

1:04:44
It didn't tell me that it does not know the person or something like that.

1:04:48
So that anything that is false, positive or negative falls under the bracket of hallucination.

1:04:53
So anyone who is, you know, in the profession of dealing with reduction of the hallucination will have
a future in this AI industry test and it is reliability.

1:05:05
Again, if you ask the same question of generative AI, you're not certain that you will get the same
response, right?

1:05:11
So how do you evaluate the responses that you're getting from generative AI?

1:05:16
So again, if I take you back to our Azure Pi Studio preview again, this is still in preview.

1:05:24
You can go ahead and explore this, but there is something called Evaluation.

1:05:30
In this evaluation tab you could have your question answer pairs.

1:05:35
Then you could search for my model and I could provide my existing data set.

1:05:44
So you could basically give a question and an answer pair and using some sort of a model, whether it
is bird or whatever, you would basically do the vector comparison of the response that is coming out
of the model based on the text, the query that you have provided.

1:05:59
And you could evaluate if the answers that you were expecting versus the answers that are coming
through from the model are you know of certain efficacy or not.

1:06:08
So for all your benchmarking needs you could use in our project we have built it custom.

1:06:13
So if you see here we are comparing 2 RFXS in sourcing space.

1:06:18
We are basically required to prepare a request for proposal documents.

1:06:21
So we are comparing the RFP that we were expecting versus the RFP that is getting generated out of
the system under the various parameters.

1:06:29
The below ones are the cosine similarity word, precise and Recon whatnot, right.

1:06:33
And we are basically saying that OK, both of them are accurate, you know or similar or dissimilar
whatnot.

1:06:39
This is in line with what Devaraj was showing in the previous nights requires large amount of training
data which is difficult to obtain and manage security.

1:06:52
So for a small company it is not, you know, possible, you know how much of investment that Open AI
has done right, millions of dollars, you know, just train those models and keep the chat jupity up and
running.

1:07:05
You can look at the figures again, the environment impact, it consumes a lot of processing and
electrical resources, you know.

1:07:13
So there is an environment impact of that extensive expensive to implement and maintain requiring
specific skills and resources which is very difficult to get need for continuous monitoring and
reinforcement learning because any model would require the reinforcement learning with human
feedback to bring up its accuracy and stuff like that.

1:07:31
And then there is fast changing environment, whatever you were learning six months back may not be
very useful or relevant, you know in the next six months, right.
1:07:41
So this is a very fast changing environment.

1:07:43
We have to keep learning in order to keep pace and stuff.

1:07:47
What are the opportunities that it is throwing?

1:07:50
So you know the mundane tasks are getting reduced.

1:07:55
You know the human requirement for those mundane tasks, probably wait for that will be reduced
because the system will be able to do a lot of those.

1:08:04
However, there are new startups and initiatives that are coming up which are powered by AI just to,
you know, give you a sense that in this particular case what this inflection 2.5 is claiming is they have a
better model.

1:08:20
So what they're saying is it is more empathetic, helpful and safe compared to even our ChatGPT and
stuff like that.

1:08:27
So I prior asking a question what is your opinion on vos?

1:08:31
It gave me a very empathetic sort of an answer.

1:08:34
But if I ask the same question to ChatGPT it probably first tell me that I I do not have personal
opinions and then it will tell you know a very practical sort of an answer.

1:08:50
So based on your need you can choose one of this.

1:08:53
But the point I was to drive home is there are multiple models which are, you know, tuned for specific
purposes.

1:08:59
Have you know are coming through you could.
1:09:03
You have the liberty to choose pick and choose.

1:09:05
And if you are in a job of the problem solving that you get to choose between these models and you
know pick the best one for your needs, you can do that.

1:09:13
Then I wanted to cover about the GPU versus LPU scenario.

1:09:16
So NVIDIA is providing GPUs making a lot of money.

1:09:19
Then there are companies like Grok, Grok with the cube which are saying that Grok is our mission to
set the standard for generative and I helping helping real time AI application from the live.

1:09:32
So the major cell point here is the speed of all these things.

1:09:35
So I can't give that AID.

1:09:36
GPI.

1:09:39
Let me ask a very quick small question.

1:09:43
Give me 100 metals from adiotic able.

1:09:51
It will ask the same question to chargeability.

1:09:57
Start this one first and probably I'll start this one and see how fast this is.

1:10:05
It is 559.

1:10:06
I think chargeability is still forming its response.
1:10:10
So you see there is a big investment happening in this area.

1:10:13
The opportunities are there increasing job satisfaction.

1:10:16
Obviously you are not required to do mundane jobs.

1:10:18
You can focus on the actual problem statements and you know derive satisfaction out of it.

1:10:24
Cost savings and efficiency, time to market, basically radiation to you know, productization is very fast
with you know generative BI helping you do all this creative processes even though you know
automatic use mundane tasks and stuff like that.

1:10:39
Personalized experiences.

1:10:41
Again think for a while Game of Thrones if if I could redesign Game of Thrones using Sora model, you
know based on your preferences and you know your likings and disliking and stuff like that.

1:10:54
Many people didn't like the last episode.

1:10:55
What if I could redesign that so based on everyone's preference so people will get hooked to that you
know series so I could provide all this personal experience eventually UX using multi model scenarios.

1:11:06
We just discussed about the text to speech, you know it needs to text and all these things then
responsible EI if you are involved in, you know this particular feed wherein EI does not always care
about the response that it is generating, whether it is harmful or stuff like that right.

1:11:24
So if you are in this field of maintaining the responsible responsibilities of this particular EI then
probably you will have a future in that on this particular thing again I had a demo probably will skip
for timing.

1:11:43
The last thing is threats.

1:11:45
Basically it could leak your personal and sensitive data because anyone and everyone can ask a variety
of questions and it you know there is no without a proper legal and process based you know
implications.

1:12:02
We will not be able to reduce the chances of copyright infringements, personal safety and stuff like
that.

1:12:09
Defects.

1:12:09
You must have seen videos you know even Moody was complaining about, he saw a video of himself
singing and stuff like that.

1:12:16
So defects are possible with the open generative coming into picture biases and discriminatory
distance making.

1:12:23
Again, the model is trained using a huge amount of data with a lot of hyperparameters and stuff like
that.

1:12:29
They are trying their best to remove the bias and stuff like that.

1:12:32
But again, the model is as good as the data that it is trained with, right?

1:12:36
So it can have come with its own biases and stuff like that.

1:12:39
So which may lead to cost, legal and reputational risks, chances of weak points or vulnerabilities in
your code.

1:12:47
So if you are using and your Visual Studio Code Copilot to generate codes and stuff like that, you
always need to verify and see that if it is not generating anything which is 101 and unwanted and
harmful content, again it would be you know.

1:13:04
So again Azure is providing A vector experience on that.

1:13:08
So if you look here, there is something called content filters, right?

1:13:23
So we could always filter out the harmful contents and stuff like that using this particular tab, so
Violence gate, whatever, right?

1:13:30
You could filter out all these things based on this.

1:13:33
Yeah, things mean the end of my slides I guess, because you want to add anything, I don't know.

1:13:44
Thanks Asad.

1:13:45
I think we are pretty much at the end, so I can stop the recording and if anybody has any question
please feel free to ask or offline we can also respond.