0:00

You also might have

0:02

heard about MongoDB in certain ways.

0:04

This is a thing that's around

0:06

and I always heard about it

0:08

and yeah, what is it? And

0:11

it turns out a

0:13

database system that is quite important

0:15

for AI to work. So

0:17

I was curious what MongoDB actually

0:19

does. And I talked to Richmond

0:21

Alake, he is developer advocate at

0:23

MongoDB. And he explained a lot

0:25

about the three essentials of AI,

0:27

which is compute, algorithm and data.

0:30

And this last thing, data, the

0:32

important thing is how to access

0:35

the data and MongoDB helps. So.

0:38

Welcome to another episode of the

0:40

beginners guide to AI. It's Dietmar

0:42

from Argo Berlin at the microphone

0:44

and I can talk a lot,

0:46

but let's just give the microphone

0:48

to Richmond and see what he

0:50

has to say about how to

0:52

make data an essential part of

0:54

your AI. Let's go! Yeah,

1:02

today we have Richmond Alakev

1:05

from MongoDB. And before

1:07

I talk too much about him, people

1:09

know the drill. I give a microphone

1:11

to him because he can say much

1:13

more about himself. And actually, first of

1:15

all, Richmond, welcome to the podcast. And

1:18

my first question would

1:20

be, why AI? What

1:23

did bring you to AI? Yeah,

1:26

that's a very good question.

1:29

Dietmar, thank you for having me on

1:31

this podcast. I'm hoping we have a

1:33

very good conversation and speak to your

1:35

audience about AI, agentic systems

1:37

and what we're seeing today, but

1:39

more importantly, MongoDB's relevance

1:41

in the current state of AI

1:43

and in the future of

1:45

AI. So to

1:47

answer your question, why

1:49

AI? What brought me to

1:52

AI? So I thought In

1:54

my career journey, it was a

1:56

natural progression. So my

1:58

undergrad was in software engineering.

2:01

After my undergrad, I

2:04

then became a software

2:06

developer. I became a

2:08

web developer building websites,

2:10

building mobile applications. Then

2:12

I got bored. So

2:15

I thought to myself, what's

2:17

a good challenge that where

2:19

I can use my existing

2:21

skills and build upon it.

2:24

And AI was the next

2:26

level, was the next

2:29

logical choice. So I tried

2:31

to teach myself AI. I

2:34

wouldn't say I tried to

2:36

teach myself machine learning, but

2:38

I realized how difficult it

2:40

was at the time and

2:42

I scared myself into back

2:44

to university. So I got

2:46

a master's in computer vision,

2:48

deep learning and robotics. So

2:51

I guess that's what we call AI today. So

2:53

masters in AI. And yeah,

2:55

that is really why

2:57

AI won because it's

2:59

a solid challenge to

3:01

it's going to be

3:03

probably humanity lost in

3:05

last invention. And

3:07

it would do it right. And it's

3:09

it's relevant, very relevant.

3:12

Yeah. That's great. We'll

3:14

come to the topic of the

3:16

last invention later, but let's

3:18

go really simple, because I don't

3:20

do software, unfortunately, but

3:22

machine learning. I mean, this is

3:25

like the machines learn like the brain,

3:27

but how in detail does that

3:29

work? Can you go a little bit

3:31

deeper in that? you

4:53

And for this you need special chips.

4:56

everybody talks about them, I guess.

4:58

Well, you Nvidia chips,

5:01

but we didn't start at NVIDIA

5:03

chips, right? Machine

5:05

learning a field or

5:07

as a general conquest

5:09

for man, it's something

5:11

that we've been doing

5:13

since I guess as

5:16

a field since the

5:18

50s, 1950s. And then

5:20

GPU became very relevant

5:22

to the conversation in

5:24

in 2012, when, I'm

5:26

going to butcher surname,

5:28

I think Alex Tversky

5:30

and his released AlexNet.

5:33

they I'm

5:36

sure I'm getting the name of the, of

5:39

the network, of the

5:41

name of the convolutional neural

5:43

network properly. If

5:45

If not, then we correct

5:47

him. I it's AlexNet.

5:49

it was a neural

5:52

network that essentially could,

5:55

that was trained using GPU, which

5:57

means that the neural, the

5:59

parameters and the

6:02

weights and biases, the

6:04

information, the knowledge

6:06

of this neural network

6:08

was actually placed

6:10

on GPUs. instead of

6:12

CPUs or wherever

6:15

processing system was used

6:17

at the time,

6:19

because GPUs had the

6:21

ability to compute

6:23

and computer mathematical information

6:25

in parallel. It

6:28

made things a lot quicker. So

6:30

it made the training of these systems a lot

6:32

quicker, and then you can then do it at

6:34

what at the time was scale. And

6:38

by scale, I mean you can make this neural

6:40

networks bigger. I

6:42

won't go too much

6:44

into the weeds, but the

6:46

long story short is

6:49

GPU was one of the

6:51

key aspects of the

6:53

journey of AI that really

6:55

accelerated a lot of

6:57

things that we're seeing today.

7:01

And there's three commodities in AI,

7:03

compute, model, and

7:05

data. So GPU is in that,

7:07

isn't that computing? When you

7:09

solve for compute, you

7:11

can actually move a lot

7:13

quicker on that innovation

7:15

journey. Yeah, yeah, yeah,

7:17

this is, it makes sense. The more compute

7:20

you have, and this is the race for

7:22

data centers and stuff like that, and the

7:24

other parts come behind. That's

7:26

actually directly a way to talk

7:28

about MongoDB, because data is the

7:31

next step. The computer is nice,

7:33

but you have to do something

7:35

with it. And you are actually

7:37

the ones that provide the infrastructure

7:39

for accessing data as far as

7:41

I, with my half

7:43

knowledge, would say, could

7:46

you explain what you actually do? Data

7:50

is very crucial. Again,

7:52

one of the key commodities

7:54

within AI that makes AI

7:56

work. And you're absolutely

7:58

correct. MongoDB is at the center

8:00

of this. We've been at the

8:02

center of this. MongoDB is very

8:04

relevant. And MongoDB is

8:07

a general purpose database that

8:09

is feature rich. And

8:11

what it does is it

8:13

stores and retrieves. data.

8:15

It allows you to build

8:17

an application that can

8:19

store data and retrieve data.

8:21

That's what MongoDB does,

8:23

but there are different ways

8:25

you can store data

8:27

and there are different ways

8:29

you can retrieve data.

8:31

We've seen that evolve over

8:33

time, but the unique

8:35

aspect of MongoDB has been

8:37

and still is the

8:39

way we structure and format

8:41

data, then store it

8:43

within our systems is in

8:45

a format that is

8:48

different to what database we're

8:50

initially thought of, which

8:52

is tables, relational

8:54

model. MongoDB has the

8:56

document data model. And not to

8:58

go into the weeds of

9:00

things, the TLDR

9:03

is MongoDB stores data

9:05

in the way that

9:07

developers think. And

9:09

that is something called

9:12

JSON. This is

9:14

a JavaScript objects

9:16

notation. It's a data

9:18

structure that has a key

9:21

value pair format. And

9:23

this is common. You tell

9:25

any developer, Jason, they understand

9:27

what that is. And even

9:29

some non -developers understand what

9:31

Jason is. And it doesn't

9:33

matter what programming language, what

9:35

application type they're using, Jason

9:38

is the most common

9:40

inter -exchange data format within

9:42

the application landscape. But

9:44

wait, we have AI. That

9:46

has not changed. And

9:48

one thing that we saw

9:50

was this LLMs, large

9:53

language models, actually had

9:55

a natural affinity

9:57

for understanding JSON, which

10:00

means that not only does

10:02

MongoDB actually allow developers to build

10:04

application and store application in

10:06

the way that they already think,

10:08

but It applies, we also

10:10

allow you to store information in

10:13

the way that LLM already

10:15

think as well. So it's

10:17

just very natural, very natural

10:19

where we are and where

10:21

we're going. That sounds like

10:23

you build an AI and

10:25

that is like structured a

10:27

little bit like the human

10:29

brain and coincidentally, MongoDB works

10:31

like with the human brain

10:33

and then you can also,

10:35

AI can also work with

10:38

it. That's a really, really

10:40

practical thing. Yeah. Yeah. It's

10:42

basically, again, I'll go very,

10:44

very a bit technical, but

10:46

not too technical. One

10:48

thing is this LLMS, they gave out

10:50

a lot of a

10:52

lot of responses, right?

10:55

And we loved that

10:57

when they first came up,

10:59

but then we started

11:01

to realize that we required

11:03

structure in the outputs

11:05

of this LLM because it

11:08

allowed us to make

11:10

what was very probabilistic become

11:12

more deterministic, which means

11:14

that we can actually steer

11:16

this LLMs to produce

11:18

outputs that are predictable. And

11:21

the best way that

11:23

we know how to do

11:25

it for developers, for

11:27

application developers, for architects of

11:29

the future is through

11:31

JSON. So now there are

11:33

modes that are called

11:36

JSON mode, or maybe structured

11:38

outputs, which puts this

11:40

LLM in the state of

11:42

providing outputs that are

11:44

JSON formatted or in the

11:46

JSON schema. And MongoDB

11:48

stores that data perfectly. I

11:51

was always trying to figure out

11:53

the connection because you have a LLM,

11:55

it's basically language output, but then

11:57

you have kind of a switch and

11:59

you can let it program or

12:01

work statistically. And I think

12:03

then this is the other mode that

12:05

comes in. Oh yeah, yeah, okay. And

12:08

then there's another thing

12:10

is just so this structure

12:12

of MongoDB supports that. the

12:15

basics thing is what's so important

12:18

about data then? Why do I

12:20

need to store data? Yeah,

12:23

well, one thing is you

12:25

can answer that question in

12:27

any point in time and

12:29

the answer would relatively remain

12:31

the same. It doesn't matter

12:33

if it was the big

12:35

data error or maybe the

12:37

computer vision error or this

12:39

generative AI error. The question

12:42

is, Why do I need

12:44

to store data? One,

12:46

you need to store data

12:48

to allow for things within

12:51

the application, such as personalization.

12:54

That's one thing that is very

12:56

understandable to most people. When

12:58

you use an application, you

13:00

want to be able to

13:02

feel that the application was built

13:04

for you and is able

13:06

to meet your needs. And the

13:09

way that application developers have

13:11

done that or have done that

13:13

in the past is to

13:15

be able to collect data on

13:17

either the user or provide

13:19

data relevant to the user, which

13:22

requires you to store the

13:24

data somewhere. Now, let's take

13:26

that into the age of

13:28

generative AI that we have now.

13:30

There's a lot of data

13:32

on the internet. but why do

13:34

we need to store data?

13:37

These LLMs can provide an output,

13:39

but they need relevant data,

13:41

domain specific data that allows them

13:43

and their output to be

13:45

personalized to you. And that's

13:47

where we come in. We

13:49

give you the, we are the

13:51

data layer where you can

13:54

actually store that data and we

13:56

give you different methods of

13:58

retrieval where you can bring the

14:00

data to the LLM to

14:02

create that personalized experience for your

14:04

customers and restore different type

14:06

of data on structured and structured

14:08

data and one of the

14:11

common type of data that is

14:13

very popular today is something

14:15

called vector vector data or vector

14:17

embeddings which is basically a

14:19

data object of like a music

14:21

or a numerical representation of

14:23

a music or an image that

14:25

is that is stored and

14:27

used for things such as semantic

14:30

search. MongoDB stores that as

14:32

well. That is why

14:34

data. We can go into

14:36

the details, but that's generically why

14:38

data. Yeah. Actually, this vector

14:40

database, I heard a lot about

14:42

this is new thing, the

14:44

new shiny toy, because it's better

14:46

to store data like this

14:48

and you have things like retrieval,

14:50

augmented generation, where you use

14:52

those vector databases. Is

14:55

it? Why is it

14:57

better? Can you say that in simple

14:59

words? Why is the vector? I mean

15:01

vector might have had that in school,

15:03

but it's like long time ago. Yeah. You

15:06

hit the no on the

15:08

head, which is you've heard of

15:10

vector a long time ago.

15:12

Vector is nothing new, right? The

15:15

way we use it or

15:17

the high fidelity of the

15:20

vector data that are being

15:22

generated might be might be

15:24

new now. We can capture

15:26

more information in this numerical

15:28

representation of data objects. So

15:30

a vector data for your

15:33

listeners is let's imagine I

15:35

have an image. I

15:37

can pass this image through

15:39

what we call an embedding

15:41

model, which is a machine

15:43

learning model that has been

15:45

trained to understand some of

15:47

the patterns within any input.

15:49

It could be image or

15:51

text, and you have a

15:53

multimodal embedding model, then provide

15:55

a numerical output that captures

15:57

the patterns or the features

15:59

of this image. With

16:02

that numerical output, you can

16:04

actually do something which is interesting.

16:07

You can take a text, let's

16:09

say the text I

16:11

am searching for a

16:13

red image, then pass

16:15

that through the same

16:17

embedding model, then do

16:19

what we call vector

16:21

search. And that

16:24

embedding model that embedding of

16:26

vector data of the

16:28

text, which will be your

16:30

prompt, will within

16:32

this high -dimensional space be

16:34

closely related in distance

16:36

to an image or

16:38

images of maybe something

16:40

with that image of

16:42

the color red. That

16:44

is the whole premise

16:47

of vector search and

16:49

embedding. Now, it's

16:51

not new, it's been,

16:53

it's had before, but

16:55

What's new is the way

16:57

we're searching and the

16:59

fidelity of the embedded model

17:01

output. Vector databases

17:03

is something that are emerging. We've

17:06

seen some examples of vector

17:08

databases, and people are adopting

17:10

this. But what we see

17:12

is nothing has changed to the

17:14

way we store and retrieve

17:16

data. The general purpose of data,

17:18

which means that MongoDB is

17:20

even more relevant than ever. which

17:22

means that we're not just

17:24

subscribing to one way of storing

17:26

data. We can

17:28

store different types of data and

17:31

provide you the needs and means

17:33

to retrieve and store those data

17:35

effectively, including vector data. So with

17:37

some database, you get the ability

17:39

to store just vector data. But

17:41

with MongoDB, you have the general

17:43

purpose database where that is optimized

17:46

for AI workloads, including the storage

17:48

and retrieval of vector data. That

17:50

is a massive information, but happy

17:52

to deep dive wherever you want

17:54

to. No, that's great because actually

17:56

I now get what it means

17:59

and the so if I have

18:01

a prompt and I want to

18:03

get something and I mean it

18:05

might be a general output but

18:07

now I have like what you

18:09

said with a red picture I

18:11

want to I don't know create

18:14

a picture that that is connected

18:16

to my pictures that I have

18:18

in my firm and now it

18:20

can retrieve all pictures that are

18:22

similar because they are somehow red

18:24

and generate something and maybe have

18:27

a style for my firm and

18:29

then as suddenly I have the

18:31

style. with those red pictures and

18:33

what I create now is not

18:35

a general picture, but would be

18:37

possible, but one that bases on

18:39

my own stuff and with vectorization

18:42

the model can find it. Did

18:44

I explain that correctly? Very

18:46

good high level. explanation of

18:48

it. And then MongoDB is

18:51

that database where you store

18:53

that vector data and have

18:55

the mechanisms to retrieve vector

18:57

data efficiently. And we

18:59

store other type of data because this

19:01

is what a lot of people are

19:04

starting to realize. You

19:06

need you need different type

19:08

of data to make AI

19:10

work, not just vector data.

19:12

So I talked to a lot of customers

19:15

that are building the future. They're

19:17

building an AI application, they're

19:19

building an agentic system, and

19:21

they start to realize that

19:23

they don't just want to

19:25

do vector search, but they

19:27

want to do. full tech

19:29

search, together with vector search,

19:31

to increase the accuracy, to

19:33

make the outputs of the

19:35

LLM more relevant to the

19:37

user's query, and to make

19:39

the entire system that they're

19:41

building more reliable and scalable.

19:44

So you don't just go

19:46

out there looking for vector

19:48

database. You go looking

19:50

for MongoDB that is

19:52

that general purpose database. So

19:54

which allows, which you know,

19:56

you're building that a good AI

19:59

application, you're building a very

20:01

solid stack and you have a

20:03

trusted data layer that is

20:05

relevant regardless of what new data

20:07

types emerge. Oh

20:09

yeah, yeah, but you talked about the customers.

20:12

I'm really interested in which areas,

20:14

is there certain areas where the

20:16

people come from or it's just different

20:19

industries, everything, or

20:22

do you have some where more

20:24

people come now for developing AI? Yeah,

20:26

MongoDB is in a privileged

20:28

position where we get to see

20:31

a lot of the excitement

20:33

that is happening in AI today.

20:35

That's because... need data and

20:37

you need somewhere to store that

20:39

data regardless of what industry

20:42

you're serving or what customers your

20:44

application is building. So

20:46

the short answer to

20:48

that question is everyone is

20:50

playing with AI in

20:52

some sense, which means we

20:54

get to work with

20:56

people in the manufacturing industry,

20:59

in the telecommunication industry,

21:01

in the healthcare industry. it

21:04

really is a very privileged

21:06

position and they trust us to

21:08

be that data layer and

21:10

where they can build the future

21:12

of their applications and the

21:15

application or different application. So

21:17

that's the straight answer to that

21:19

question. It's everyone. Crazy.

21:22

And I read on

21:24

the website every month

21:26

there's 175 ,000 new

21:28

developers starting with MongoDB

21:30

and this is like But

21:35

this is just the new

21:37

people. How many people

21:39

work with MongoDB? Can one

21:41

say something like that?

21:44

I can't give you an

21:46

exact number, but I

21:48

can give you an understanding

21:50

or perspective, which is

21:52

MongoDB is one of the

21:54

most popular databases today. There's

21:58

a lot of people

22:00

that are familiar with MongoDB.

22:02

Most developers are. So

22:04

that's the skill at which

22:06

we are relevant and

22:08

which we have an opportunity

22:11

to help developers and

22:13

businesses build this long lasting

22:15

applications within this AI

22:17

era. So long story short,

22:19

it's hard to put a number but

22:21

we have There are

22:23

hundreds of thousands of developers

22:25

that are aware of MongoDB and

22:27

there are thousands of businesses

22:29

that trust MongoDB as that data

22:32

layout for their applications. Yeah,

22:34

that's a good answer because I

22:36

knew MongoDB as a name. I didn't

22:38

know the advantages and you told

22:40

me something about this, but yeah, this

22:42

is a brand name that everybody

22:44

knows somehow. You go a little bit

22:46

into programming and then you come

22:49

there and see your people. It's

22:52

a good branding. It's

22:54

good branding, but also the

22:56

technology itself is very useful. I'll

22:58

tell you a personal experience

23:00

of mine, which is when I

23:02

was in university, I

23:05

said earlier, my degree was

23:07

in software engineering, but I

23:09

don't tell people this. I

23:11

hated coding because to build

23:13

an application, you have to

23:15

learn the front end, the

23:17

back end, the database and

23:19

they were just different worlds.

23:22

In fact, in the

23:24

web application space back in the

23:26

day, we used to have three different

23:28

people. developers building one

23:30

application or more, at least

23:32

free, because you needed someone that

23:34

knew HTML, CSS, and JavaScript. They

23:37

needed someone that understood the

23:39

backend. It could be Java

23:41

or wherever backend language. They

23:43

needed a database administrator that

23:45

understood your database. But this

23:48

is what MongoDB did for

23:50

me. It unified the application

23:52

stack. which means with the

23:54

mindset of Jason and thinking

23:56

about Jason, it made it

23:59

easy for me to just

24:01

understand how things connect through

24:03

the application stack, the front

24:05

end, the back end to

24:08

the data layer. And

24:10

it was so significant

24:12

that it changed my career.

24:14

So I became a

24:16

full stack web developer because

24:18

I understood MongoDB. I

24:21

embraced that JSON, that object -orientated

24:23

programming mindset, that JSON mindset,

24:25

and used it all across the

24:27

stack. So we saw the

24:29

emergence of full stack web developer.

24:31

You probably have heard of

24:33

different stacks like a mean stack,

24:36

main stack. And those

24:38

are things that MongoDB helped within

24:40

the application space. But we're

24:42

not stopping there because that same

24:44

thing we did for people

24:46

like myself when I was a

24:48

bit younger is what we

24:50

are doing for the AI engineers,

24:52

for the AI application developer

24:55

today in the world of

24:57

generative AI. It's the same

24:59

aha moment. Oh,

25:01

yeah. So that is something

25:03

you didn't want to program. Actually,

25:05

do you still program or is

25:07

it AI or do you just

25:09

manage your agents and their program? Well,

25:12

I wish I had a bunch

25:14

of agents that were coding for me.

25:16

That will allow me to do

25:19

more podcasts and talk to

25:21

you more. But yes, I still

25:23

code on a daily basis. I'm

25:27

in a unique position

25:29

where I enjoy building things

25:31

and having a very

25:33

good understanding of what's going

25:35

on in the application

25:37

development landscape and then communicating

25:39

that information to developers.

25:41

to our customers. We do

25:44

that in MongoDB in

25:46

different ways. We have an

25:48

educational platform. We have

25:50

a developer platform where you can

25:52

see different articles. And we

25:54

actually speak to our customers one -to

25:56

-one and their specific AI team. So

25:58

long story short, I

26:00

code more than ever. So

26:02

it's not yet there. You said at

26:04

the start, AI may be our last

26:06

invention we make. Can

26:12

you go more for programming

26:14

or in general, like your

26:16

vision for that? I'm

26:18

curious to hear that. One

26:21

thing is, the

26:24

promise of AI has

26:26

always been to replicate

26:28

human intelligence. The question

26:30

is why, right?

26:32

Why have we wanted to do

26:34

this? And it's because we can get

26:36

a bit philosophical here. Great,

26:39

yeah. Life can

26:41

be laborious, which is you're coming,

26:43

you're born into life, you're expected to

26:45

go into work, you work for

26:47

several amount of years, then you retire

26:49

at age 60 something, and then

26:51

you get to enjoy the money. It's,

26:54

for lack of

26:56

a better word, that

26:59

there is more to life than

27:01

that, but... We need a way or

27:03

we need some form of entity

27:05

that can help us with this laborious

27:07

part of life that can then

27:09

allow us to experience the creative part

27:11

of life or aspects of life

27:13

have not seen. I

27:16

think there is a

27:18

future where AI allows us

27:20

to be very productive. It

27:23

allows horses as humanity to

27:25

explore different areas of that.

27:28

of the experience of what

27:30

it means to be humans

27:32

that we don't have the

27:34

opportunity or the scope to

27:36

explore today. It

27:39

can be

27:41

humans' last

27:43

invention if we do it right

27:45

and also if we do it

27:47

wrong. It

27:50

goes both ways. But I'm

27:52

not going to bring the

27:54

tumourism vibe. I think AI

27:56

has a lot of benefit

27:58

to what we do today

28:00

as humans and it can

28:02

improve a lot of productivity

28:04

for a lot of people

28:06

within different domains. I

28:09

actually have a typical question here in

28:11

the interview and that would be how probable

28:13

do you think is a terminator or

28:15

matrix scenario so a little bit of do

28:17

more is my one to get. I

28:20

think there is there is

28:22

always a potential for certain outcomes

28:24

right it would be. it

28:27

would not be intellectually honest

28:29

to say that there isn't

28:31

a certain percentage for that

28:33

outcome happening. Does

28:36

it happen exactly

28:38

like Terminator? I

28:40

doubt it. It

28:42

might be less cinematic

28:44

or dramatic. I

28:47

think any technology

28:49

in the hands of

28:51

the wrong people

28:53

can cause harms. And

28:57

that's the same for AI. That's

29:00

the same for the internet.

29:02

That's the same for any technology

29:04

we've ever invented. Yeah.

29:07

Yeah. No, it's like, this

29:09

is a probability. No, but

29:11

it's a possibility. Yes. Yeah. That

29:14

is good. Yeah. One thing

29:17

about yourself. Do you

29:19

use chat GPT or for what?

29:21

What things do you use AI

29:23

programming? Obviously, probably, but do

29:26

you have other things that are

29:28

funny, interesting or helpful? I

29:31

use chat GPT. Again,

29:33

I need to have a very

29:35

strong opinions of the things that

29:38

are coming out of LLM. So

29:40

I use chat GPT. I use

29:42

Claude. I use Gemini. I use

29:44

pretty much the whole chat interface

29:46

that are coming out. I

29:49

use it for different things, right?

29:51

For my side projects when I'm

29:53

programming, I use it for thinking

29:55

as well, for debating, for looking

29:57

at other perspectives of some opinions.

30:00

I use it for pretty

30:02

much a good amount of

30:04

what I do today. But

30:07

something I'm realizing that tomorrow

30:09

is that there is an

30:11

aspect of when I use

30:13

this LLMs, where I begin

30:15

to over engineer the problem. I've

30:20

never seen a computational

30:22

entity or even an

30:24

intelligent entity think about

30:26

2 plus 2 so

30:28

deeply as much as

30:30

an NLM will. But

30:34

there are times where I

30:36

see myself over -engineering when I'm

30:38

working with NLMs when the answer

30:40

is actually quite simple. But

30:44

yeah, I don't see a question. I use

30:46

it for a lot of things. And there's

30:48

some things that I don't think I can

30:50

use it for yet. But

30:53

it's really interesting, this point

30:55

where one gets tempted to...

30:57

ask everything to the machine,

30:59

even if it's simple, like

31:01

2 plus 2, use a

31:03

calculator. But the people

31:05

then asking those questions, those huge

31:07

machines. Great, yeah, that's a good

31:09

thing. And what can come out

31:11

of this is paralysis through analysis. If

31:14

you analyze too much, you

31:16

don't do. I think that's

31:18

it's over -engineering. And that

31:20

happens regardless of an LLM

31:22

or Rivaan LLM, which is

31:24

One thing that a lot

31:26

of people are going to

31:28

realize is LLMs and the

31:30

error of AI and the

31:32

AI models, they amplify a

31:34

lot of things, right? Productivity,

31:36

but they can also amplify

31:38

a lot of things that

31:40

we don't want them to

31:42

as well. So we have

31:44

to use them with, not

31:46

with caution, but use them

31:48

with habit. I find that

31:50

it's better for me to

31:52

have an objective. that I

31:55

want to get when I'm working with

31:57

Vela Lambs and then maybe time box

31:59

myself because you could go into the

32:01

rabbit hole. Definitely

32:03

those long chats. Yeah, great. Yeah,

32:05

yeah. And then now, now starting with

32:07

voice mode and so you're even more

32:10

tempted. Yeah, one has to control their

32:12

own use. Yeah, it does. And one

32:14

thing is One

32:16

thing I use LLMs for this

32:18

new interface is like you mentioned

32:20

voice mode is I use it,

32:22

I use it for learning different

32:24

languages as well. So I'm learning

32:26

a certain language at the moment

32:28

in time. I have a human

32:30

tutor, but at the

32:33

same time, sometimes I go

32:35

into chat GPT, I put it

32:37

in voice mode and I

32:39

say, Speaking Japanese,

32:41

let's speak in Japanese now and

32:43

we go back and forth in

32:45

Japanese very terribly on my part.

32:47

Excellent on the chat GPT part. And

32:50

I use it as a tutor

32:52

as well. But again, it's not replaced

32:54

my human tutor. I

32:56

totally get that because for

32:58

myself the listeners may know I'm

33:00

married to a Cuban wife

33:03

and so Spanish is a choice.

33:05

I learn languages with also with the

33:07

voice mode and it's really, yeah. This

33:10

is something so interesting that you

33:12

now suddenly can go there and talk

33:15

in a language and it helps

33:17

you and it understands your problems and

33:19

can correct you. Really, really great. Japanese.

33:22

Oh, yeah. But that's a

33:24

harder Spanish. It's quite easy

33:26

compared, I think. Japanese. Whoa.

33:30

Yeah. Okay. Yeah.

33:33

Long ago, I did some, some

33:35

candle like this. fighting

33:37

thing and they had like terms

33:40

and I try to remember them

33:42

and it was like not easy

33:44

but great it's like yeah cool,

33:46

cool challenge. I'm terrible now at

33:48

my Japanese and I'm just doing

33:50

it as it's fun right and

33:53

this is what I was talking

33:55

about the human experience. If

33:57

we have more time

33:59

to do certain things. What would

34:01

you do? For me, I

34:03

would learn some of the languages

34:05

of cultures that I'm interested

34:07

in. I

34:10

would love to do that full

34:12

time, but we have this laborious

34:14

aspect of work that then gives

34:16

us the permission to do the

34:18

things that we actually want to

34:20

do. But I'm in a

34:22

fortunate position where I made a conscious

34:24

decision to go into AI because I

34:26

wanted to do it and I was

34:28

good at it, I understood it, but

34:30

now I want to learn Japanese as

34:33

well. Yes,

34:35

it's actually, this is really like

34:37

this philosophical thing is a good

34:39

thing to end on because what

34:41

do you want to really, really

34:43

want to do? And what should

34:45

you do if there's no work

34:47

anymore? Many people will have

34:49

to think about what are they

34:51

have to have to develop

34:53

hobbies or passions and so on.

34:56

I think

34:58

there is

35:00

a future. there

35:04

is a possible future

35:06

where the word work disappears.

35:11

Wow. Yeah, great. Yeah, I

35:13

love it. I

35:15

think we have to stop now

35:17

because you can't top that

35:19

because this is really something. I

35:22

mean, yeah, the work is

35:24

done and by machines, by

35:26

thinking machines, they are great.

35:28

Yeah, but Richmond, tell

35:31

us where we can find

35:33

you, where can we find

35:35

MongoDB, and where can we

35:37

find some information? We put everything in the

35:39

show notes, but Richmond, tell me. Yeah,

35:41

we put everything in the show notes. You

35:43

can find me on LinkedIn,

35:46

right? Type up my

35:48

name, Richmond Alake, and hopefully there isn't

35:50

an imposter on LLM trying to

35:52

impersonate me, but I will show up.

35:54

and you can find me on

35:56

LinkedIn, reach me on LinkedIn. Well,

35:59

you can also see some of the stuff, good

36:01

stuff I'm doing over at MongoDB. We'll

36:03

put some stuff, some link

36:05

on the show notes, especially a

36:07

piece that we use to

36:09

explain AI stat and agentic system

36:11

and AI agents. I'll put

36:13

some articles that will help your

36:16

listeners understand what's going on

36:18

without going too much into the

36:20

details. And

36:22

yes, find me on LinkedIn, connect with

36:24

me, reach out to me and

36:26

let's learn together. Perfect,

36:28

perfect. So, Richmond, I really

36:30

thank you for this. And I also

36:32

thank you that you explained things not

36:35

on a deep tech level where I

36:37

was like blanking out here. I understood

36:39

what you said and was great having

36:41

you. Thank you very much

36:43

and look forward to speaking to you

36:45

soon. Yeah, we do that. Thank you.

36:48

Yeah, another thing I learned, JSON,

36:50

I always heard that, but that

36:52

is the language that makes it

36:54

easy to access data. Okay,

36:56

more I don't need to know,

36:58

but now I know there's

37:00

a thing that all programmers like

37:02

and that makes the communication

37:04

between LLMs and data easier. great

37:07

for us. And it seems

37:09

to be quite successful. 175 ,000

37:11

new developers each month. That's quite

37:13

a number. Yeah, so I

37:15

hope you, like me, learned a

37:17

lot about neural networks today

37:19

and machine learning and had fun

37:21

listening to Richmond. And the

37:23

last thing, don't forget to subscribe

37:25

to the newsletter. and

37:30

also follow us on your podcast

37:32

app. We would be happy to have

37:35

you there and in the next

37:37

episode as well. So signing off, Dietmar

37:39

from our Godot Berlin.