Well, hello everybody. Welcome back to DayToday with Greg Michaelelsson. Uh, I'm real excited to be here joined by

0:16

Roy Zalta, uh, whom I just only met recently actually through Pod Stars, right?

0:22

Yeah, that's correct. Uh, a great community actually. I uh I been, you

0:27

know, meeting with great people from that community and it's a great way to connect, hosting a podcast and learn

0:32

more about, you know, AI and other stuff as well. Awesome. Well, give us the uh the

0:38

rundown on you. What uh what's your background? What are you interested in? What what are you doing these days? Well, it's actually pretty interesting.

0:45

Uh this week was my first week in Microsoft. Actually, I started a new job there.

0:51

Okay. Um, but today I will talk uh, you know, because I'm pretty much new over

0:56

there, I will talk with my hat as the chief officer in PWC Israel actually.

1:02

And I think it's really interesting, you know, talking about the state of AI and I'm talking about the state of AI this

1:08

week, you know, because it's changing so constantly. I mean this week let's says

1:15

uh I think the hot topics right now are mold bot probably the things people are

1:20

mold bot have you heard about that I mean clawbot mold bot oh yes yes yes

1:26

yeah so I think this is the hot topic I mean this week and you know the thing I

1:31

do every morning is spending almost like an hour on Twitter uh ro and getting so many information

1:38

about what's new in AI right now because this is I'm not sure it's the best resource course, but it's probably the

1:44

resource that you'll find like the most interesting and know unusual uh stuff online about what's new in AI and what

1:50

people are actually doing. And I have really good cool uh good use cases to

1:56

talk with you about. I mean uh stuff that I heard people are doing and I'll be happy to share with you guys some of

2:02

that. Yeah. How has how's this all this whole LLM thing impacted the way that PWC did

2:08

business or does business? Well, it's uh I mean PWC Israel specifically PWC next

2:15

actually the company I worked for uh they decided as same as other companies like big corporation companies decided

2:22

they want to be agent companies actually which is you know pretty much big words for saying uh we understand that agentic

2:29

AI it's it's great it's still on the early days but we understand that also

2:35

uh in the upcoming years uh companies will have you know human human employees

2:41

and identity employees and AI employees and they're going to work together so that that going to do and some other

2:47

stuff human going to do and I mean it's really good idea to actually prepare yourself for that and think about what

2:53

can I automate already on my company how do I prepare that and you know also

2:58

identity I mean how do I treat those uh aentity employees I saw the uh CEO I think of McKenzie

3:06

came out and said that they've got 20,000 agentic employees right now and uh

3:11

I thought uh yeah that doesn't seem right like maybe that's hype. I don't know what do you think

3:17

that that's actually a lot. I mean I know the PWC the global company is

3:23

really advanced actually on whatever related to AI they have actually something called agent OS which they

3:29

implemented for many companies like big big banks in the US which is like their way of like do you know NAT so it's like

3:36

the corporate way of NAT how do you can create automatic workflows on your company connect various tools and you

3:44

know internal tools and this is something they're really invested on and they think it's really uh going impact

3:49

and I know they have like customers that are using it and I think you know it's

3:55

funny because we're talking about PWC everybody thinking about financial company but the advisory of PWC it's

4:01

huge I mean they understood like on the early days that you know we're sitting with those like you know most of the

4:08

companies in the world right now like you know and so we have this resource and we must use it to you know do

4:13

advisory and then from advisory you know uh cloud computing became so popular. So

4:19

they decide okay we're going to uh we do we're going to do professional services of cloud computing cyber security and

4:26

most recently of course AI and right now it's actually agentic AI they understand that if you can implement agentic uh

4:32

agentic work folks with whichever uh cloud provider you want you can do amazing things uh for businesses and

4:39

actually been seeing it myself I mean we've been doing amazing stuff for Israeli companies Israeli government

4:46

it's been really popular I I I'm curious as to how you feel about how much of it

4:51

is hype versus how much of it is legit. Are are organizations actually getting value out of this or is it mostly a

4:56

science project? Uh I'm pretty much afraid of the air bubble. I think everybody are is

5:04

probably know as you know I think somebody told me that people thought

5:09

that I mean they treated uh cloud computing as a bubble as well uh back in those you know those days probably like

5:16

10 years ago but I mean over here I think the hype is mostly because you know it's everybody are talking about it

5:23

it's really hype and the the phase of the development is really like it's fast

5:28

I mean but if you look at that uh entropic CEO you know he says like almost a year ago AI is going to write

5:36

most of the code and you know everybody you know thought that sounds ridiculous but at the end of it if I'm looking

5:42

today and I'm pretty much you know talking with so many companies and developers everybody using uh I mean uh

5:49

cloud code or other tools I I can say about myself personally I haven't

5:55

written like myself a lot for more than a year I mean I'm not writing code And

6:01

from what I'm understanding, big companies and technology companies, uh the companies don't expect you to write

6:08

code anymore. I mean, if you're writing code yourself, you you're too slow. You won't you won't make it today. So, I

6:15

mean, uh devel developers are the the big getting the really big impact from

6:21

agentic AI. what we're talking about like you know the agentic uh uh workforce and effect

6:27

we're talking about this is when it when it comes to like financial institution and you know those uh things you're

6:34

going to uh uh encounter on your day-to-day life that's when it's going to be like really big impact because

6:40

developers are always the early adopters of those technologies well certainly at Zerve we found huge

6:47

adoption of agentic coding so we're Zerve is a coding environment we we do Python and R and SQL coding. It's like a

6:53

agentic development environment for data. Uh and we discovered when we

6:58

rolled out our agent that 95% of the code that gets written in Zerve is written by the agent. And so and that's

7:05

that's legit. Like that's real data. That's not hype or anything like that. People are going out and they're

7:10

discovering that using the agent to help you write code is outrageously more efficient. Uh so

7:17

I do think you're right that developers are are early adopters.

7:22

what's what's the who's next in terms of adopting this stuff? Um I'm going to

7:27

tell him that I think uh in my opinion the big impact going to be on the CFO

7:33

level and you know the financial uh employees in my opinion it's going to really big make a big impact. My wife

7:41

actually she's economist so I pretty much know this uh field from uh you know

7:46

closed look and I think there's so many stuff they do there. I mean from data

7:52

anal an analysis to you know uh doing a lot of many repetitive uh annually uh

7:59

things they do all the time that can be automated with agenti but you know

8:05

what's the challenge you know why they not already using it because it requires a lot of business logic and when it

8:11

comes to business logic you need really good context engineering and in my opinion it's only comes when you

8:17

understand the business well and the business logic And it it's really hard because you know I was surprised when I

8:24

became a developer that most of the uh the there's no not many documentation about developing. I mean the developer

8:31

world is lack of documentation. So the financial world is more lack of documentation. There's so there's really

8:38

a a big problem of documentation. you know people left people are leaving people are going and it's not an easy

8:44

things to do and there's so many you know uh uh financial workflows that are

8:50

broken for years and they keep be being broken so if you add a genti to a broken

8:56

process it's just going to keep being a broken process you know only works so yeah so what I found where I uh

9:05

advise to companies is that we're you know first trying to understand the business logic and sometimes we

9:12

understand that the business logic is broken that's and the way working it's working at the

9:19

moment so agent I won't help here you need something else so it's really important to understand the difference

9:24

between you know automation and uh agentic workflow so they comes they come

9:31

together but sometimes automation is enough and sometimes agentic without

9:36

human uh interpretation and human uh you a consent inside of it can be really

9:42

it's not it's not the great fit in my opinion. Yeah. I've never seen a aentic workflow

9:47

that didn't that could work that didn't have a human in the loop like we're we're not there yet.

9:52

We're not I I mean for developers as well and I don't think there's any field that human in the loop can be out of it

10:00

right now. That's the that's one of the challenges for sure.

10:05

Yeah, the guardrail problem is a real problem. uh an agent is only as useful as the things that it can do. But as

10:10

soon as you let it start doing things, then suddenly it can do things and it can do things you don't want. And so

10:16

it's uh it can get scary. Yeah, for sure.

10:21

So you said you had some use cases that you were interested in talking about. Tell us uh what uh what you have in

10:26

mind. So yeah um I'll start by saying to the people that are listening to us uh I

10:34

almost a year ago start develop the first aentic uh multi- aent uh process

10:40

uh in my career which was end up as being one of the first uh multi- aent

10:46

system uh in Hebrew in Israel actually I found out it was for a big insurance company here in Israel and uh it was a

10:55

big success actually but There was so many challenges because it was early. I

11:00

mean it's still early but you know controlling a bunch of agents together answering uh customers about problems

11:08

going to databases doing rag and all those kind of stuff together. It's really hard to evaluate as well. So I

11:16

found myself that eval are really big challenge but also a really great

11:21

opportunity for product managers and that's where I you know end up uh uh finding about the thing called the

11:28

product generalist actually which is like a new profession that's saying you know uh product managers needs to be

11:33

better at AI but they can really enhance and make uh AI project faster if they be

11:40

really good at you know uh the evaluation part because evaluation really requires is understanding the

11:45

business and building a good data set for evaluation for agents. So if it's a

11:50

explain a little bit more about what you mean by evaluation. Yeah. So if we're building a genetic

11:57

system either one agent or multiple agent or not agent at all just you know

12:02

a simple rod we want to have uh a set of questions and answer that we expect the

12:08

agent to answer uh based on those uh uh questions. So let's says we're doing a

12:13

rag application and drag application is about uh answering about documents. So

12:19

we need to take those documents and create enough questions for the uh LLM to answer about and we can evaluate it

12:27

about the uh questions in two ways. The first way is human evaluation where we

12:33

ask him about a bunch of questions. we get the answer and then the business will you know evaluate each uh answer

12:40

either it's good or bad and mostly we go for binary so good or bad and they can

12:45

add comments about it. That's uh pretty much always the first step because we're

12:50

doing not a big a large set of uh questions and we want to get like some kind of like a big picture where we act

12:57

and when we go for scale we want to do automation and we want to have a CI/CD about it then we go for uh uh auto eval

13:06

and on auto eval we do like you know big evaluation and many questions and we get

13:11

we must set the uh expected answer which is called grand truth And every time you

13:17

know the answer the agent will answer about each questions another LLM will evaluate the answers and you know he

13:24

will score how much this uh answer is close to the ground truth uh and then we

13:30

get like a score mostly and the score will be like you know a range from I

13:35

don't know 0 to 100 or from 0 to 10 and we get this score and it will tell us

13:41

you know how good it get us some kind of like a metric how good is our agent right now and from there we can grow and

13:48

can set a KPI. We want to be at I don't know 98% and we won't release a new

13:55

agent unless it's you know above I don't know a 95%. So it's really becoming you

14:01

know more like a machine learning classic machine learning to uh produce those kind of agent and you know the

14:08

connection with the business is really important here and the business they get their metrics they know how good uh the

14:14

application is and they can they feel they feel safer about releasing it to customers.

14:20

So the first step in evaluating these guys is human intervention. You could dive in there, read the answers, score them. Uh and the second

14:28

at scale using a large language model to do that evaluation and scoring. Yeah. Is that a training step or is that

14:34

happening in real time or or is that how does that work logistically? uh we want to do it as many as many

14:41

times as possible because we want to know how each change affect our uh application in our LLM and we want to

14:49

make sure that you know if I added the prompt a little bit or I I added an extra tool to the agent I want to make

14:54

sure I didn't broke anything else. So I want to run the same evaluation over and over and I want to make sure the

15:00

evaluation is updated with the data. Let's say I added another tool. I want

15:06

to evaluate this tool and I want to make uh and ask questions that will evaluate the ability of the agents to call the

15:13

tool use the tool and get you know the actual data and answer uh uh do actually

15:19

uh the retrieval from the this data. So you want to do it as mighty as possible you know on classic uh programming we

15:26

have unit testing uh on machine learning we have smoke testing and here we have

15:31

evil actually so it's a must and it's really it's an art that I think there

15:37

are some kind of really cool people I'm following online that are becoming an artist of evaluation and you know it's

15:44

it will become standard and as AI adoption will rise uh eval will be the

15:50

key in my opinion Now it is I guess there's two kinds of adoption, right? So you have companies

15:56

whether they're going to adopt like these chat bots that are smart and agentic that can work with customers and

16:01

answer their questions and do things for them. But then there's also customers are are customers going to be willing to

16:06

talk to an AI in order to get stuff done. Uh, and right now I try to bypass

16:11

those things just personally. Like if I see that I'm talking to an automated chatbot, I'm I'm like representative

16:18

representative representative and I try to bypass those things. Do you do that? So So I actually tried to go with it. I

16:25

mean there's really uh there's a really great company here in Israel called Wonderful.

16:31

Wonderful. Yeah. They're called Wonderful AI and I really I talked with them a couple of

16:36

times. They're amazing in my opinion and at the moment they control the entire

16:42

customer uh uh agents LLM agents of all of many big uh companies of Israel. They

16:49

uh phone call agents. So I let's say uh I had a how how do I find out about

16:55

them? I called uh uh Makabi which is like our uh you know uh customer support

17:00

for uh healthcare stuff and the agent answer for me and you know Hebrew is a

17:06

really hard language to to get a good uh tune and the find the uh how the how

17:13

words are spoken in Hebrew and the Asian spoke amazing and it answers about my problem in a really good way. And then

17:21

when I had really you know complicated uh questions, it uh it took uh it took

17:27

the uh the decision to move my uh uh appointment to you know a human

17:32

representative but it answers successfully and it was really amazing and that's the reason this company is so

17:38

really amazingly uh doing right now and all the company's banks are using it right now because they have a fine tuned

17:44

Hebrew voice agent and they really amazing I mean really I and a really

17:50

good combination. These are text based. These are actual voice conversations over the telephone with a AI.

17:56

Yeah. And it's a real time voice agents. And not only that, they can I I know they're using tool under the hood and

18:03

stuff like that. And they do it really really good. I mean it's they do it in real time. There's no like delay or

18:09

something like that. It feels really natural. And I mean I'm a technologist so I like to talk with LLM sometimes.

18:15

But I understand you that you're saying I don't talk with an a with an LM. Give me a real human. I need the answer. But

18:21

they're doing it really good and the escalation is proper. I mean really good. Uh I mean

18:27

my claps to this company. It's really they're really good. That's fantastic. Cross the AI the verbal the vote the the

18:35

spoken AI agent with customers. I mean yeah there are some companies who

18:41

do it really good. Uh I understand the difficult the it's really difficult for you humans to get okay it feels like the

18:49

LM won't help really help me by uh inquir is too complicated but I think

18:55

this is the field where we're going to see LLM's really controlled it stills the customer uh relationship

19:04

the thing that that bothers me about those kinds of interactions is when the agent gets it wrong and they start

19:11

telling me stuff it's not right. And then if that happens, you know, then not only do you have to chase it down,

19:17

but maybe you've also made bad choices based on what the agent originally suggested. And I've used coding agents

19:24

enough to know that they sometimes go down a rabbit trail and sometimes make some bad choices about what, you know,

19:29

like what packages to use or what implementations to try or what coding structures uh to utilize. And so like if

19:35

I'm calling up, you know, the airline and I'm trying to change my ticket, I need to be sure that that ticket is like

19:42

accurate and like, you know, that the times are right and that the seat is right and all that stuff. And I don't

19:48

know, I'm still a little wary about trusting the agent to to put me in the right seat on the right plane and everything.

19:54

I understand the concern. And you know, I mentioned earlier the health uh customer agent and the I mean, you know,

20:01

I bet the company that implemented this solution thought we're going to implement it and give people health

20:07

advices. This is even more dangerous. There's risk for every kind of those use

20:13

cases. But I think the risk of not doing it is probably higher for those companies,

20:19

for the companies. But what about for the users? What do you think? Um

20:25

sorry in my opinion it's it's going to be harder for you know

20:31

people which are not tax savvy to uh you know believe those uh solutions. I mean I believe it most of the times and you

20:38

know I'm not a health uh expert and not a flight expert so it might be wrong. I mean I don't know any I didn't encounter

20:46

any power any airline AI solution. It's interesting actually to see it but it

20:52

will be really sucked if it get wrong you know and I get on the wrong flight or something like that but in my opinion

20:59

I actually cancelled my flight uh like a week ago or so and I tried to talk with

21:05

the customer support and it was human customer support and you know they did in my opinion I'm I'm sorry a really bad

21:12

job and it was really hard for me to you know try to you know reschedu this flight and I thought I thought to myself

21:19

here like in this scenario I would prefer to talk to an LLM. I mean the human here the human here did not help

21:26

me. So, in my opinion, an LLM, you know, because I they asked me about what specific other dates I would like to

21:32

reschedule my flight, and I had to tell him every time, and he had to check if it's possible. I mean, for an LLM, I

21:37

could just tell him, would I don't know, give me the the cheapest uh flight on uh

21:43

I don't know, on March, and he probably could go and do that or tell me I can't do that. You know, with humans, it's

21:50

harder. I mean, especially when I'm not an English native speaker and you know,

21:55

uh, sometimes our customer support is not in Israel. So, yeah, for sure.

22:00

Yeah. And I guess we just have to sort of learn the boundaries of what these things can do, right? So, like like if

22:06

I'm talking to S, do you have an iPhone or are you an Android? iPhone. Yeah, my man. So, like if you're talking to Siri, you

22:12

know that you can ask her to set an alarm or start a timer or turn up your brightness or turn down the volume. But

22:18

if you asked her to delete an email account or to you know uh filter your

22:24

emails or you know do there's certain stuff that you know she can do and stuff that you know she can't do. Uh but the

22:30

important thing is that she knows what she can do and can't do. And so if I ask Siri to you know like uh I don't know

22:37

delete the last text thread with uh with Roy then she knows that that's not

22:42

something that she's capable of doing. I don't know if the LLM's know that. like they're so eager to please.

22:49

Oh my god, I'm so excited for uh Siri to get smart so so soon with the Gemini

22:56

ideal because it's something I've been waiting for. I mean, I saw the demo for Apple Intelligence and where you know it

23:03

notify you that your grandma is uh on that flight and you need to go and pick it up and stuff like that. And I thought

23:09

it's I mean if if someone could put it up it will be Apple and then you know Siri is still not as uh as good as we

23:17

expected but I think you know as I mean I'm I'm an Apple Apple fan sorry I think

23:23

as every other thing they do just you know give it a time and I think it's it's probably going to be amazing right

23:29

now when when cereal gets smart and you know if if you have the Apple Apple

23:34

ecosystem it could do like amazing stuff. I mean I have cameras on my house

23:40

connected to my home kit. I have uh I have everything I mean connected other than my cats. I think everything else

23:47

here is connected to Apple ecosystem. So the integration together can do like amazing stuff and automation and you

23:53

know and the thing about the ecosystem is closed. So if you wanted to feed your

24:00

iMes uh some agent as a part of software, that's not something that's an option.

24:05

Like that's not possible to do unless you're Apple. Uh the they've got an opportunity I think just like Android

24:11

does uh to harness all that they know about you. And the good thing about Apple is that a lot of the compute is

24:18

done on device. Uh that the edgebased uh LLM computing stuff is really smart.

24:26

Uh, and so getting smaller models that can run on your phone, I think, solves a

24:31

lot of privacy issues, too. I bought this. I have the iPhone 17 Pro Max and I bought it because I thought uh

24:38

it could be the only model that could run like edge SLM, small language models

24:44

or something like that. Hopefully it will come and then you know they probably will use I don't know in my uh

24:51

for what I what I wrote uh they're going to use like a small version of Gemini on

24:57

phone for some of the stuff and it's going to be routing and if it's complicated they go to the Apple cloud

25:03

and it's going to be secure in the same way Apple secure everything and boy it's

25:08

going to be amazing it could be like that's it I mean that's the game changer I mean it's probably going to affect uh

25:15

chat GPT customers for sure. I mean it's going to really affect them because you know uh Apple users are tend to be

25:22

really tax saving. If it's going to be really good they going to ditch it and probably you know if it's going to be

25:28

pay for Apple intelligence people will do it. I will do it for sure and well here's the thing. Uh, I think

25:35

there's going to be a bunch of antitrust stuff that goes on because if Apple signs up with Gemini and that's the only

25:42

large language model that you can use to inter integrate with your operating system, that's an antitrust issue,

25:48

right? Just like it was with uh Safari and Chrome on the phone. Like now

25:53

there's like a whole mechanism for switching your default browser and switching your default search engine. But when the iPhone first came out, none

26:00

of that existed. Like your default was your default. And then there were lawsuits and blah blah blah. And I think

26:05

all that same stuff is going to happen. It is going to happen. But, you know, we'll see how it goes because maybe I

26:13

don't know. They're going to they're going to have their developer event and they're going to have the like app

26:19

developers do their uh integration with Apple intelligence with this new world. And we'll see how it goes because not

26:25

everybody is going to use Gemini. Maybe they're going to use I mean even the model they going to use it's not it's

26:31

going to be Gemini Gemini with change of model weights so it's not really going

26:36

to be Gemini you know it's under the hood Gemini but really great move yeah

26:42

and I think I mean I'm not an Android expert but I think uh Google probably I

26:48

mean going to have their own thing and you know I saw how Gemini is integrated into uh Google Home actually and it's

26:55

really amazing too I tried And you know, you can actually ask about, you know, uh, smart

27:01

stuff on your smart home like you can ask Gemini to turn off the lights and, you know, do all those automation on

27:07

your house. It's pretty amazing. Yeah. Know, I have cameras around my house. I use the Nest one, so it home.

27:13

And every evening I get the home brief where it tells me everything that happened at my house, which I think is amazing. My wife hates it. She's like,

27:19

"Would you stop with the home brief already?" But yeah, that tells me that, you know, a deer walked by and there were some chickens in the yard and

27:25

somebody rode by on a bicycle. I think that stuff's fascinating. It just highlights the key things around the

27:30

house, like if UPS came, that sort of thing, uh, to deliver a package. Uh, you know, that that integration is

27:37

definitely cool. But it's more um sort of like observational and less sort of

27:42

active in terms of like, oh, you know, you weren't home and so I locked your doors for you kind of thing. Uh, which

27:49

you can set up those automations. Those are those are just like rule based but they're not uh agent based if you know

27:54

what I mean. Well uh my cat toilet they have an automated toilet and it's you know they

28:00

have a AI on it. They have a camera in it. It recognize which cat was on the

28:06

bathroom did what and I get like an element that I can ask about it. I just saw it released and like you know I saw

28:12

it. It's the company called Toya. Really really good company actually. It's it's great product. T O Y A.

28:20

T U Y A. Yeah. U Y A. Yeah, they have they have smart home as

28:26

well, so you connect all your smart devices over it. But the the smart cat toilets, it's really I mean

28:32

such a Yeah, it's such a niche. You know how many people like their cat to have a

28:38

smart to cat toilet? But you know, that's that's wild. Yeah. No, people who

28:45

have cats tend to really, really love their cats. And I've met a fair number of cat people that would love nothing

28:50

better than to monitor their cat's digestive health. I have my uh Zatim robot. It's the

28:58

dreamy model. It can it has a mode where it can, you know, scan my house and look

29:03

for my tags. Not clue. Just look for them and take a picture. Go find uh Rusty, huh?

29:10

Yeah. No, I did. I tried the vacuum thing, but

29:16

uh I have to keep the house too clean for the vacuums to work. So, it's kind of an irony, right? So, I would love

29:22

where they would just go, but like if I leave um a charging cable, you know, like if there's a charging cable for my

29:28

phone that's like laying out on the ground, then the vacuum eats it and and dies and then I have to take the vacuum

29:33

apart and and so on. So there's new smart ones that actually know uh you know see obstacles and you

29:40

know some of them even robot rock have like an arm that can move it.

29:45

Shut up. Yeah, the I'm I'm a pretty I'm a vacuum expert. It's the robots by the way.

29:53

What's the best one? Uh I have I have the Dreamy and I really like them because you know the camera and you know

30:00

that looks for my cat but I was really hoping for a vacuum that you know it's

30:05

part of the Apple ecosystem and you know everything works together and you know probably because that's the thing about

30:12

smart houses and you know the possibility of AI with it if it was integrated and there was you know one

30:19

ecosystem because I have so many devices from different companies that would be great you

30:24

Yeah, Google definitely has done better with I with the home Google Home than Apple did with HomeKit.

30:31

Uh, for sure. Unless unless something has changed in the last two or three years. I switched from Apple to Google uh probably two

30:37

years ago. And it is significantly more reliable just in terms of connectivity and and speed and and that sort of

30:43

thing. So, Apple's got some catching up to do on that front, I think. Yeah, for sure.

30:49

All right. So, what's the next big thing? What what did you join Microsoft for? What what do you do there? I'm going to uh do a job head there. I'm

30:57

going to work with startups uh on the AI and that data uh subject of matters. So

31:04

I'm going to work with uh customers uh same that I did PWC but I'm going to

31:09

work with them on the Azure uh part of the world. soon got to work with um

31:14

Azure have Microsoft have Microsoft fabric which is like the data platform at Azure have so many you know the

31:21

reason I really joined Microsoft because you know I follow all the cloud companies I saw what they do and I think

31:28

the vision Microsoft have about aentic AI and how it's going to you know be part of the organization it's really

31:34

like the the greatest vision I saw and I thought to myself well I got to be a part of it I love it I I really like the

31:42

vision and I I'm really excited to be there. I mean, you know, people are amazing. The vision is amazing and it's

31:48

it's a big step. I love it. It's the way they treat the jet AI and know I think

31:54

it's the only company that see right now that agents going to be employees and

32:00

you know that you know to treat them as employees and think further about that. Fascinating. I'm still a little

32:06

skeptical. So thinking about the hyperscalers um Azure has fabric and Google's got um

32:15

vertex right and uh you know you've got SageMaker and a bunch of others have you

32:20

evaluate like compared them which ones do you prefer the best just from a hyperscaler perspective um it really depends on the use case I

32:28

work on people I see with many uh uh clients uh each of them use you know

32:33

different cloud providers so I had to use their you know uh cloud providing and stuff. I think all the companies

32:39

they each one of them solved you know different uh uh problems they had like

32:44

IWS solved different problems Google solved really well the data uh world and

32:50

you know they're really good with agents and they have ADK which is a really good

32:56

agent platform in my opinion framework Microsoft by the way have a Microsoft agent framework which is also open

33:03

source and AWS have agent core which is really good uh product too so they have each of them have really good

33:09

solutions and it's always depends on the customer needs and what they have right

33:15

now. I mean if the clients works with GCP you go for GCP and you see what there but it's really good to you know

33:21

learn from it but you know many clients work uh hybrid they have some of it on

33:26

GCP some of it on others and as a as a developer it's really fun you know to

33:32

work with because you you get to know everything from everything and there's so many really good good use cases and

33:41

uh when I worked here in PWC I had really good use cases with government uh

33:46

uh uh offices. Uh they had really good use cases of agents and LLMs and I

33:53

really see the impact of it because I already see those uh uh chat bots and

33:58

LMS on production and I see how it helps people to you know find information about their uh you know government uh uh

34:06

department really easily. Yeah. No, that's good. They do that. The large language models are great for

34:12

search and for uncovering information and you distilling a lot of stuff down into into exactly what you need to hear.

34:18

What do you think things are going to look like in five years? I've got a really skeptical kind of point of view on what large language models are doing

34:24

to people. Uh, you know, I keep see I keep I saw a congressional hearing not

34:29

long ago about how things like Tik Tok and and all the other social media are so bad like they have such bad effects

34:36

on people and particularly teenagers uh that are using them. Do you It seems to

34:42

me that the large language models are just as bad uh if not worse in a lot of ways. Uh are you optimistic or

34:48

pessimistic? Do you do you agree that what are what are the big pitfalls and what's coming in the next five years? I

34:54

think LLMs will bring uh you know cures and medicines. I really believe that. I

35:00

think it's it's something that going to happen. They're going to discover new materials and it's it really going to

35:05

help and benefit uh human uh lifespan and human uh uh life quality. But it is

35:12

comes with a price and I think the price is the uh uh uh the mind and how it's

35:18

going to affect our mind and you know our uh because social media you know as much as I love social media and I love

35:25

LinkedIn it is it it has its cost and I think with LLMs it will solve you know

35:32

some of people's loneliness maybe but as much as it will it will uh make people

35:37

not even looking for human interaction because it you know it solves their loniness. illness and it is going to

35:42

affect you know let's say for school I don't know how kids I don't have kids yet but how kids are go how kids are

35:48

learning today I mean why why would you learn today if you have like all information in the chatbot I mean how

35:55

how do you teach as a teacher that's something I'm I'm not very uh an expert

36:00

on the education and how people uh teacher are teaching today with AI but I mean it's it sounds for me like a

36:07

problem why would I listen to a teacher where I can just you know have an LLM and why would I even learn why would I

36:13

even learn how to code if I if I'm people are not coding anymore you know I mean what's going on I mean it it has as

36:21

much as the problems that it it solves I mean I I'm really scared about um a wall

36:28

where uh LLMs are going to uh I don't know I I believe I'm I'm still saying

36:35

thank you to Chad I think it will remembers my thank you and appreciate and I'm I'm

36:43

polite, you know. So, you're treating it nicely for when it's our overlord. I mean, I'm trying.

36:51

I don't hate that. But yeah, that's to me it's like, you're completely right. Most teachers that I've come across are

36:57

just sticking their heads in the sand on this and saying, "Don't use it. You can't use it. It's cheating." Blah blah blah. But at the same time, it's a tool.

37:05

And I honestly don't know what the world is going to be like when, you know, like

37:10

once Elon Musk released his robot, releases his robots, which I will buy on day one. Uh, you know, and everybody's

37:16

got a household robot to do your laundry and clean your toilets and make your bed and stuff. Uh, you know, and you're I

37:23

don't know. I think Twitter and and LinkedIn and all these are going to become a wasteland of of bot generated

37:29

content if they're not already. uh you know when when Elon Musk bought Twitter

37:34

somebody estimated like 95% of the content on Twitter was generated by bots

37:40

but now it's so easy to generate content it's like what are those platforms even for uh but then of course if we stop

37:46

writing and creating new stuff then there's no new stuff to train the large language models on and they sort of

37:51

stagnate a bit I think so that's right like that's like the big problem because you know uh they always

37:57

release uh better models which trains on more data But as time goes, more and more of the

38:04

data is going to be uh synthetic and LLM generated. So LM is going to train and

38:09

study based on LLM generated content. So it probably go worse, you know, it get

38:16

worse the for the funny. I don't know how how they going to solve it. I don't know. I mean by by the way most of the

38:24

uh industry talking about that transformers is not the final. It will not be AGI if that's what people think

38:32

it there will be something else. I mean at the end of it transformers they just predict the next best token. It it

38:38

really did it well. Um and you know there's really like what's going on

38:43

right now with Mulbot and all those uh uh stuff. I mean if you guys heard about that that's crazy. Like people do crazy

38:51

stuff with it. Like they have this uh if you heard about that they have their own social media. They even have a website

38:58

when they can rent uh humans to do a job that they can do.

39:04

Roy, hold that door real quick. We're we're about done. We have like maybe five minutes to wrap up. Let me go

39:09

grab the door real quick. One second. Sorry. Yeah, no problem.

39:18

Hello. Yes. Awesome. Very good.

39:24

Yeah, just right here. Thank you.

39:35

All right. Sorry about that. They're just bringing I bought a couch and uh they're it's I've been waiting on it

39:41

forever. Now they're here delivering it and they're like, you know, it's so hard to get them here. Anyway, sorry.

39:48

Continue. I didn't mean to interrupt. Uh well I was just saying that you know

39:53

I was uh thinking about what's the what's people are talking about right now and people are mostly talking about

39:59

mold bot clubbot uh it has so many names their names change every time but mostly

40:06

they talk about it because right now it has the ability to you know connect itself to every kind of external tool.

40:13

It can even like you know without you if you allow it to do anything it can you know buy itself uh uh on Twitter on Two

40:22

phone number and then call you it can call on your behalf and it can do like amazing stuff. It's 100% not 100% not

40:31

secure which I don't recommend to run on your you know personal machine and you need to be really careful with it

40:36

because it you know it can control your entire uh uh uh computer and do whatever

40:42

it like and but it's really really amazing how it became a trend and you know the really big story right now is

40:49

how it's about mold book which how agents actually created their own social

40:54

media where they complain about their humans which is like really funny. I don't know how reliable it is, but it's

41:01

hilarious. That's hilarious. All right, give us a give us a warning or a recommendation.

41:07

So, to say there's somebody out there that has never used Chat GPT, uh knows

41:13

doesn't know very much about this AI thing and they're skeptical, uh but they are curious, too. So, what would you

41:19

give give that person some advice? Well, I think my favorite LLM is Claude

41:27

for sure. And I think a person who never touch either one of those AI tools and want to be like their jaw drop should

41:35

probably go and check out Claude and let it you know build for him stuff because

41:40

I think what Claude really defined that how uh AI become a commodity and the

41:47

thing that you can do with it. I mean you can do amazing stuff with it application and it's so smart and can do

41:53

like if it's something that related to uh uh uh generative generated stuff uh

42:00

out of nothings and not like you know reading stuff uh and and more like creating stuff. Cloud is amazing. I mean

42:07

it's such an amazing tool. They do amazing job over there and I heard uh Claude uh five is going to release uh

42:15

this week or next week. So, I'm really excited about that, too, because it's probably is the smart the smartest uh

42:22

person uh on the planet right now. It's It's amazing.

42:28

So, you're saying I need to cancel my chat GPT Pro and move over to Claude, huh?

42:33

Um, no. I'm I'm keeping my chat GPT because it has it already has my memories which I can uh export out of it

42:41

and it knows me well. So, I I need I need a memory. I need a personalization.

42:46

That's something clog and not as good as ch. Yeah, that's fair. Well, hey Roy, thank

42:53

you so much for taking the time. This has been a really fun talk to talk and uh I'm I'm excited to hear about how it

42:58

goes at Microsoft and see the cool things you do there. Thank you, Greg. It was an honor for me

43:05

to talk with you. Thanks. Appreciate that. Likewise. All right. Thanks so much. I really appreciate it.