2:40

Good morning everybody. Welcome to all of our users that have joined and folks

2:46

that uh may not be familiar with Zerve yet. Uh we've got some exciting stuff to

2:51

announce today about uh some developments that have happened here in the product. And I'm joined today by the

2:56

other co-founders of Zerve. I've got Philly, our CEO, fearless leader. Say hi.

3:01

Hello. Hello. And Jason, our our uh the big brain behind Zerve, chief technical officer.

3:08

Say hi, Jason. Uh thanks, Greg. excited to to be here.

3:13

Awesome. I'm pleased as well. So, I'm Greg Michaelelsson, chief product officer here at Zerve, and we three are

3:19

here today to announce something really cool that uh we have just released or

3:24

will be shortly releasing later today uh in Zerve. It's going to change the way that people interact with their code and

3:30

with their data and it's going to be pretty spectacular. I think I was thinking back earlier this morning to

3:37

how I got connected with these two fine Irish gentlemen uh years ago. uh Philly

3:44

uh reached out to me on LinkedIn and uh he said, "Hey," he tried to pick a fight basically uh because I'd been in a

3:50

company called data robot and we were working on automated machine learning and

3:56

we both Philly and Jason and I had all been kind of working with this problem where data scientists didn't really have

4:03

the tools that they needed to tackle the problems that they were trying to solve. and we were coming at it sort of from

4:09

different ends. Um, but when we put sort of all the all three of us together, kind of the magic happened. Um, and and

4:17

we I think we created something really spectacular that's just getting better as we continue developing. Yeah, you g you gain a lot with uh

4:24

picking a fight. So, I think it's kind of cool to get into it a little bit. It was just we it was at the time um when

4:30

data robot had had released the co forecasting engine for the US government which I just thought was this

4:36

spectacular like result but how they got there was a mystery to me um with automated machine learning and that's

4:43

when I started to pick the fight with Greg and get into the details about how it was built and um a lot of that chat

4:49

was what formed uh kind of what we're going to talk about today. I remember Greg, you were talking about like um

4:55

trying to solve that problem with Jupyter Notebooks that took an hour to run and so someone would pick up where

5:00

someone else left off and having to run that thing from start to finish took an hour just to get started if I remember

5:06

correctly. Yeah, it was a wild time. Um we were we were working on simulations for

5:12

predicting um h where to target folks for recruitment into the vaccine trials

5:19

and we were working on a simulation that was based in Jupyter notebooks. we were trying to move super fast and so uh we

5:26

didn't have time to get all the infrastructure and stuff set up. So we're just working in Jupyter notebooks and notebooks turned out to be pretty

5:33

good tools for what we were trying to do because they give you the ability to uh to to interact and and iterate really

5:39

rapidly to u write some code and then run it and then see the results and be

5:44

able to kind of really quickly iterate. But there were quite a few shortfalls, lots of shortcomings when it came to

5:51

making that actually work. One of them in particular was not really being able to collaborate uh because it's pretty

5:57

challenging to use notebooks with a team uh just because of like dependency

6:02

issues and uh code version control and like all those kinds of issues that you

6:08

see with with notebooks and you guys were seeing the same thing, right, Philly? Yeah. Yeah. It was it was wild to us. So

6:14

speaking to you was like vindication at the time. Um you know it was obviously the middle of CO as well. So Jason and I

6:21

were working on problems and we were in our separate places in rural Ireland and we're looking at these big companies

6:27

that have raised a lot of money and thinking, "Oh, we must be missing something. There's no way, you know, that these guys are working with

6:33

technology like that. They must be more advanced than we are." Um, and so from where we were sitting, it seemed like

6:40

how is this not a solved problem? like how had the world gone to Google Docs and everything else, how could you not

6:46

have something that would work for me sitting in Ireland and you sitting in Nevada that would work between us? Um it

6:52

see it seemed like it had to be a solved problem. But when you saw such a largecale project working off that those

6:59

frailties, I think that's what made me say, "Okay, Greg, you need to get on a

7:05

flight. Um I'll meet you in Ko and South Street in Dublin. We have to have a few points to talk about how to solve the

7:10

problem." And what kind of added to that then eventually was um I was lucky enough to recruit our

7:16

head of engineering um who I previously worked with, really smart guy, Mike. And I remember you guys, because I'm late to

7:24

the data world, I remember you guys explaining to Mike what how notebooks work and how you can make 1 plus 1

7:30

equals 7. And he and him basically saying there's no way that that could be how it works. There's no way that this

7:35

whole industry is based off off of that. And of course, there's exceptionally amazing things about notebooks and what

7:41

they're good for and and we get we got push back initially to say I want to do my exploration there and I'm happy that

7:48

it's messy and I'm happy that things don't work. When I moved to production, that's a separate thing and that's totally cool. People can have their own

7:54

processes, but there was still no reason why it couldn't be a solved problem. Um, and so I guess yeah, after the Pines and

8:01

South Street a few days later, we're in a a different office in Leon Street in Dublin where we we wrote the initial

8:07

parts of our new architecture um that could that could give you stability um

8:13

without you know breaking things. So so that it was impossible to go 1 plus 1 equals 7 but still have uh all the

8:21

explor exploratory capabilities you wanted. and and like yeah I reiterate it's not that you have to go okay Zerve

8:27

is about getting to production it's not about that you don't need to change your process but just to get all the benefits

8:32

and great things out of notebooks doesn't mean you should lose um lose how you might progress that project and how

8:39

it might actually make an impact and yeah it felt like a problem really worth solving that you could increase or

8:45

elevate the impact of data by doing and solving this problem and we had push back about how big a problem it was to

8:51

solve but um I'm glad I'm glad we solved it. Yeah. Uh Jason, for for the folks that

8:57

may not know what Philly's talking about with the 1 plus 1 equals 7 thing, we're super close to this, but can you kind of

9:02

break down what were the specific things that uh were were problems and how we

9:08

kind of initially overcame them to come up with Zerve as it existed yesterday? Yeah, of course. So um yeah, like um so

9:16

I guess we'll we'll take um an assumption that people know what Jupyter notebooks are. Um so it's a notebook

9:22

that you can write code and it's uh very interactive. So a youer can write some code like a couple of lines you can run

9:28

it you see the results if they look uh like you want them to look uh you can move on to the next cell which is your

9:34

next couple of lines of code that you can execute and uh effectively um while you're doing this h all of the results

9:40

that are being produced get saved in memory. So you're updating it like live.

9:46

So it's very reactive. It's really good for kind of like h seeing small changes to your code. So you can kind of see

9:52

things uh live and that's why people like it. So it's a really quick way to iterate on code. It can be very visual

9:59

because you've got your code and then you've got the results directly underneath it. Um but some of the

10:04

challenges with it uh is that because it is like interactive and you're like in full control of like when you run a cell

10:12

and there is no order to it effectively. So it doesn't run top to bottom. um you

10:18

can uh hit like real problems with uh we'll say um uh environments. So this is

10:25

like if I send one to Greg and Philly, they might have different versions of Python installed, different packages and

10:31

it mightn't get the same results or be able to run it at all. Um so that's a big problem with collaboration. And then

10:37

there's also as Philly touched on there was a issue with uh we'll say just

10:42

getting it useful for a business. So typically uh things get deployed into the cloud. So you put it into AWS,

10:48

Azours, GCP or some other kind of a data center. And to be able to do that it's

10:54

not it's like not trivial. So you need to uh kind of rewrite the code typically. So the two things we were

11:00

trying to solve for was one was collaboration and two was to get it into

11:05

production or get it into the cloud. And how we did that stability with stability. So how we did that was

11:11

we uh kind of re-engineered what happened when you press the run button on your piece of code. So uh when you

11:17

run it, we'd actually spin up some cloud resources. We'd run it, we'd cache the results, save uh some metadata around

11:24

it, and that meant that you come back to it, rerun it, and you'd always get the same results. Um and then also h

11:32

everybody could be working together at the same time. So that was kind of the first problems we were looking to solve.

11:38

And and it looked it looked like this, right? So if I could share my screen,

11:43

it looked a bit like this. Yes. And um the so it's very different I

11:48

guess from what people would be used to if they were if they were used to working with a notebook which is just uh

11:55

you know cell after cell and kind of stacked vertically. And so this is like a this is a DAG. It's like a graph of of

12:01

code. And it it runs similar to a notebook, but you start with code. You'd run it. you'd see your output and then

12:07

all that those those outputs would be passed down to the next block. And what we what we discovered is that most of

12:14

our users were were immediately jumping into full screen mode uh or code and

12:19

variables view or whatever you want to call it to be able to kind of have a more comfortable more spacious editor

12:25

and be able to preview variables and see what's coming out of of different uh uh

12:30

code that they might run. Right. And the yeah I guess the other big thing

12:36

that changed it this year as well uh was that we have the agent. So now there's

12:42

an agent inside of Zerve that can code with you. It's designed specifically for

12:47

working with data. Um so it is um very much like you have a cursor and claude

12:55

and kind of tools like that for software engineering that's uh they're engineered to work on like large repos of of code

13:01

and across multiple files. The serve agent is designed to work with uh code. So it's able to work like a data

13:08

scientist or a data analyst kind of like get insights or uh build models and all of the things that uh data professionals

13:16

would typically work on. Yeah. Yeah. So, so really like about six months ago, I guess to give a bit more

13:21

context, like we double down on where we feel the benefit of um and where agentic

13:26

technology can go. So, we think the sky is the limit. The ceiling's unbelievably high for how um for how advanced you can

13:33

make a product in this space. And what we did was look at how um how the others

13:40

h have worked it in different spaces like Jason said in in like the likes of cursor, the likes of claude code um and

13:47

take what we've learned so far in terms of doing data science development and apply that to to Zerve and what we've

13:53

seen since has just been rapid explosive growth related to that. And it's massively interesting to see how that

14:00

changes how you work with with data. And um more and more I've been referring to

14:05

that as data work because it was specifically looking at how you do modeling and all of these sorts of

14:11

things that were you know kind of usually sat inside a data team. We're seeing rapid growth outside of those. So

14:17

like what you can do um doesn't depend on the team you're in anymore. Um you

14:22

know it's super interesting to see data work um being done across the organization. Um, and it means that just

14:30

the speed to insights are just so so quick. And what we've coupled that with is then the familiarity of what experts

14:37

were used to now. So like as we're getting into what we're releasing today, we're releasing notebooks which bring that familiarity. So if you're coupling

14:46

uh a data an agent made specifically for data science with a a view um uh and an

14:53

interface that's you know proven to be the most effective way of working with data. um you get something really

14:59

powerful and that's why we're so excited to be launching launching that today and and you know getting into it today.

15:05

Yeah, there's two things going on here, right? Because the world of coding is is kind of changing pretty dramatically

15:11

with all these agents that have come in. In case anyone doesn't know, Jason, can you describe like explain what is an

15:16

agent when it comes to writing code? Uh sure. So um so chat GPT kind of like hit

15:22

the hit the scenes maybe like two three years ago and it kind of introduced like large language models. So this became

15:28

like a available to everybody and everybody used it for for chat and what they do is you give it um like a piece

15:35

of like a a prompt effectively. So you ask it a question and it's able to generate an answer. Um so this is like a

15:42

so typically people would have started with like natural language. So it's like writing essays and everything like that.

15:48

Uh but code itself is just uh like text just like um uh like any other piece of

15:54

a document that would be written and it turns out these large language models are really really good at writing code.

16:00

Um so what an agent is is effectively the application of a large language

16:05

model for the uh generation of of code and then the other kind of um innovation

16:12

around it is the context that's provided. So, um, the context window or

16:17

how many like tokens or words effectively these agents can hold when in like memory when they're dealing with

16:23

a problem has gotten like really really large. So, it can hold like hundreds of thousands of of like uh tokens or like

16:30

words at a time effectively. Um, and that means that you can fit in like lots of code or lots of information about uh

16:37

the data that you're working on or the problem you want to solve. And that means that these large language models

16:42

uh have like the full context to be able to understand the problem you're working on um to be able to like suggest really

16:49

useful code. Um and then uh typically what they'll do is you'll give it um a

16:56

prompt or an instruction natural language. Um it'll tell you what it's going to do and then it'll go and

17:02

execute it. You can see the code that it's written and then you can like approve, edit or like ask it to make

17:08

some some edits. So it's like a really is like a kind of a a co-orker with you

17:13

kind of like writing writing code and it can do it at like um phenomenal speeds

17:19

and you can have multiple ones working at the same time I guess is the other kind of like big big multiplier.

17:25

The the context is important. So last night I was making uh dinner uh and I

17:31

wanted something easy so I asked for uh some ideas for easy dinners that kids

17:36

might like. And the seventh option it gave me was Korean beef bowls. And so I

17:43

was like, "Okay, let's do number seven." Uh, and it gave me a recipe for Korean beef bowls. And then I said, "Well, I

17:49

need a shopping list." Uh, and by the way, put the shopping list in the same order as the stuff in the store so I

17:55

could just walk through the store and get it. Uh, oh, and then don't put section headings. And so I kept I kept

18:00

kind of going on building on that conversation. And what's happening behind the scenes there, just like in

18:06

Zerve, is that the context in that the agent is aware of or that the large language model sees is growing. Like at

18:13

the beginning, it didn't know I was making Korean beef bowls. Uh but as I kept asking questions, it remembered

18:19

what we talked about. Uh and Zer does the same thing, right? So it's going to be able to see the data that you have

18:24

uploaded. It's going to be able to see the the databases you're connected to, the previous code that you've written,

18:30

the other output that you've seen, just like in that Korean beef uh uh dish,

18:35

your project is like the you know that conversation that we had, right? Exactly. And the one extra thing is

18:42

probably as well that it can take actions. So it's like um that it's able to um so it's able to uh make files,

18:50

edit files, create files, write code like it's able to take the actions that

18:55

a user would with inside an app say and and for a typical data workflow. So

19:01

we obviously have usage patterns for for how uh how how these projects go. And so

19:07

taking all that into account that context is automatically in that agent. So it's um it's not like that needs to

19:13

be reexplained. You're coming in and saying, you know, uh it could be something like let's let's find let's

19:18

see is there any interesting insights in my file. Um and that's very broad to an agent that's not made specifically for

19:25

doing data work. Um but not very broad to to something like Zerve where it both has the context and then crucially the

19:32

the environment in order to be able to run that. So run the code, see the results, keep adding like you said,

19:38

adding the context like your list, Greg, but instead it's adding each set of results back in. The environment is very

19:44

important there because you're getting something that can run code, retain results, um, and and be able to build

19:49

and build that context to to get those insights super fast. So it'd be like if I had the Tesla

19:55

Optimus robot to make my Korean beef bowl. Yeah. Which, by the way, I'm going to have the

20:02

moment it's available. Just exa ex and and maybe uh the robot is

20:08

trained in in soul um on on the best uh best recipes.

20:13

All right. Excellent. So, uh we've talked a lot about it. Let's show it. Uh Jason, will you pop it up on the screen?

20:18

Let's let's look at the new interface for reserve and see uh how it works and where everything is situated and give

20:24

everybody a tour. And and while that's while you're bringing that up, Jay, looks like we've got a question in as well, which will

20:30

probably be answered through the demonstration here. So um looks like a question in that's uh from Chitel saying

20:37

how does the agent actually help me write or debug code? So I think we're going to be able to cover that through just a demonstration here too.

20:44

Yeah. So we'll get into it and we'll see the the agent kind of do a full end to end um kind of example in in a minute.

20:51

Um but this kind of just uh here's one I I met earlier I guess um for uh the new

20:57

interface. So you've got your classical kind of like notebook view, but you've

21:02

also got the the DAG view that um uh Greg showed earlier. So this is like a

21:08

visual representation of the uh of the project. Um so you've got your coding

21:14

interface that you have here. You've got your outputs in line like you'd be familiar with. Uh you've also got this

21:20

where you can explore all of the the variables uh that got created by this particular uh block. Um, and then, uh,

21:29

these two are connected, so you're able to kind of like easily navigate. Uh, you have AI generated descriptions of each

21:36

of the blocks. Whoa, whoa. We're going we're going super fast here. Wait, wait. So, can you go to the middle and just shrink it so

21:42

that your code like move your little boundary between them. Uh, so that Yeah. So, we can we can see the code a little

21:48

more. Move it to the right so that you have Well, not that far to the right. You done switched because it's so

21:55

skinny. So you have some control over how things are laid out and you've got like your output in line just like you

22:02

would in in a notebook type view. Uh and and you showed how you could kind of

22:07

click to jump to a particular cell. Uh right? Yeah. So you've you can navigate between

22:15

uh so both the the notebook and the canvas are connected. H so you're able

22:20

to use it for for navigation purposes. Now, this is one main difference, one major difference here between what we're

22:27

looking at here and a notebook is that uh you you've got this nonlinear thing

22:32

happening, right? Uh yes. Yeah. So, here we've got the start of the project and then uh what we

22:38

have and what is this what is this doing here? This is reading in a CSV file. So, this is reading in a CSV. So, this

22:44

is just a data it's one of our quick start examples. So, it's a data salaries

22:50

uh file that has just a a load of like historical values for different job descriptions across the the US.

22:58

And looking at the DAG, it looks like you're starting with that on the left, but then the data frame that you create

23:04

when you read in that CSV file is being fed to lots and lots of different cells. Uh yes. Yeah. So, the DAG kind of

23:11

illustrates that dependency. Exactly. Yeah. So, each of these connections do two things. So one is uh

23:17

for data inheritance. So you pick up data from uh upstream blocks and then

23:23

the other one is for orchestration. So if we were to do like a run all this first block would run and because of the

23:29

way that we've set up the uh the compute and we know all the dependencies all of these ones in parallel can run at the

23:35

same time. H so that's kind of unique to to Zerve in in that sense. And then each

23:40

of it uh of the downstream blocks can can run h and you have full kind of

23:46

reproducibility of the results as well. Got it. So the output from each block is

23:51

going to show inside the blocks on the right and you can get a closer more spacious view of the code over on the

23:57

left side. Exactly. And you can toggle on and off. You can go full screen on the notebook

24:03

view and you can also go full screen on the uh canvas view if you want to just uh do like an exploration of it. Um and

24:10

then we've got Oh, I see the description up there. Yes. How did you get to that? Show one more time. Uh so there's kind of two new controls.

24:17

Uh one is uh the block details. So this is an AI generated description of each of the blocks. Um so every block you

24:25

click on has an AI generated description. uh just so if you're kind of new into a a project uh you can kind

24:32

of just easily navigate it. So uh say I shared this with you uh you might come to this view and start to click and

24:39

instead of having to try to figure out just from the the code and the results what's happening you can read this uh

24:44

description and then the other thing as well is that there's the uh the gallery. So a lot of the time it's the images

24:51

that you're kind of like interested in. So you can kind of quickly come here, you can look at it, read a description

24:56

of like what the image is and then you can go to the go to the cell if you want. So it's just a kind of a easy way

25:03

to navigate across all of your different kind of like pieces in your in your project. Got it. That's fascinating. And now uh

25:11

on the far right side, what's going on over there? Uh over here, this is our agent. Um so

25:16

what I'll do is I'll jump into a different uh just to start from scratch. So this is your empty state. Um, so what

25:22

we'll do just to show to answer the question from um I don't know how to use the agent code.

25:28

We're still seeing the old one. Have you jumped into a new one yet? Oh, sorry. Might need to stop and start sharing again. Has it switched to the new?

25:34

Oh, there it is. There it is. We see it now. Um, so what we'll do is we have a couple of quick start options. So we connect to

25:41

a database. H, if we wanted, we could upload a file. Uh, we can import an

25:46

existing notebook and it'll like pre-populate it. uh or we can start from like a an exist kind of a quick start

25:52

data set. So I'll pick something who that's kind of like small so it can get started. So it's added to the file so

25:58

it's available to the project and then there's also this uh generated first uh first block. Um so this is the agent on

26:07

the right. Um so what we're able to do is with chat we're able to we'll say just from a coding perspective what

26:13

we're able to this this is Titanic. Yeah. H so this is the Titanic data set. So, so, so everybody's familiar with this,

26:18

right? This is this is one row in this data set for uh each passenger on the Titanic and it tells things like how

26:26

much did they pay for their ticket and were they male or female and details about the passenger and then whether they lived and died or lived or died at

26:33

in the in the wreck. Yeah. Exactly. Yeah. So then, so every every data scientist who's ever

26:39

been on Kaggle has seen this data set. 100%. So then um so what we'll ask it to

26:44

do is uh to do some EDA and just to spice things up a little bit, we'll ask it to to show us things we've never seen

26:50

before with some uh uh cool charts. Um so what it'll do is it'll take this

26:56

prompt and it'll start to think it'll look at the the data. So it's able to

27:01

load the files to get the context. It's able to look at the Oh, so it's actually going to be able to see the the output of this of the first

27:09

block. Yes. Yeah. So that's how it figures out um what it is. And it can even start from you can start without the first

27:15

block. Um it can start from scratch and the first thing it'll do it will be loaded and every time a new block is

27:21

created with a new set of results it'll update its own context and it's able to kind of like one block at a time just

27:27

like a data scientist or a data analyst or anybody who codes would. It kind of

27:33

uh uncovers kind of like uh results one piece at a time updates what it'll do in

27:38

real time as a a as a result tool set. So this

27:44

Yeah, go ahead. I was just going to ask you about that. Oh yeah. So here it presented a plan. So this plan has like load and inspect uh

27:51

kind of like a survival pattern. So kind of give like an outline of all of the different things it'll do. And I'll just

27:58

click follow agent here. So as the agent is working and creating blocks, it'll start to navigate across the um the the

28:06

different kind of things that it creates and we can see we get like a live progress. So uh it'll kick off all of

28:12

these like sequentially. Um if it has a chance to run two of them in parallel, it will. Um and then it gives an update

28:19

of like how its progress is going. And then you can also see like the reasoning so each of the steps that it does um

28:26

throughout. Um, if you want to be working while the agent is working, you can do that too. So, you can click on a

28:32

block and uh, effectively it won't move you off. You can edit it. You can create a new block uh, in between. Um, so here

28:40

I've just created this one. I can be working here while the agent is working on other other parts of the project.

28:45

And you can rewire them. Yeah. If you block you just created, this isn't just visual. So you can use

28:50

it to not only like rewire but also to like create and like um h so you're in

28:56

full full control of the uh the DAG. So it's not just a visual representation of

29:02

the the project but you can pick like dependencies and build your build your DAG uh interactively.

29:08

And look at that speed. Oh my god. Scroll down Jay to see the new stuff being created. It is just wild. Like

29:15

it's we're we're live as Greg would say we're living in the future. Yeah. And you can start to see kind of

29:21

the charts that it's producing. It's got the AI generated descriptions. Um,

29:26

what do they call these? Are those violin charts? Is that what those That's a violin chart. Yes. Yeah. I'd call it a fiddle chart.

29:34

I'm a little cont you you you play the fiddle, Greg. Yeah. Only when uh when forced.

29:42

Um, so yeah, we can we can judge at the end whether or not it's like producing kind of like stuff that we've never seen

29:48

about this data set before, but it seems to be giving it a a good goal. And I get like what's what's

29:54

unbelievably cool here from a demo perspective, you might pick up on it is like you're not touching the mouse here, Jay. Like you can be to be moving

30:00

around, but all of this is what's happening in front of you. The agent is doing everything here. Yes. Yeah. So it's like just on

30:07

autopilot. If I click on a if I click on a block and I wanted to like say work on

30:12

this one, we'll see that the follow agent message shows up. So now I'm not in the follow mode and I can be working

30:19

on this block. And you can also get the agent now inside a block with the ask AI. Uh so you can get it to make like a

30:26

block level edits. Uh so wait, explain exactly the difference there between what's happening on the

30:31

right side and what's happening in that block right there. Yeah. So this is kind of um so we would

30:37

have talked about context a lot and this is kind of like a narrowed context. So the changes so when I'm working over

30:42

here in the chat it's at the level of the entire project. I can like at at

30:49

like a block or something like that. So very quickly I'll just add this and I can go like at inspect or whatever. I

30:56

can give it the context of the block and I can ask it to make edits. It'll present a plan. Uh, but when I'm in this

31:01

ask AI mode in a block, uh, it's kind of like more direct for kind of like faster

31:07

edits to the the code, we'll say. So, I could ask it to do something like, uh,

31:13

update the comments uh, to be funny. It doesn't have to be,

31:18

it can be like real, but it's like just an example. It could be anything. It's got the full context of all of the other

31:25

blocks. So, it's got everything available to it. it can go and read outputs from other blocks if it needs to. And here it is. It's added something

31:32

about Sherlock Holmes and things like that. So, and and Jason, it might be worth talking a little bit about like what we're

31:38

seeing in terms of people singleshotting projects versus like that functionality

31:43

where you're looking to go and do bit by bit or iteratively build your project. Like h how much are we seeing people

31:49

move to the lovable type model versus the cursor type model? And just for more context, what that is is where someone

31:56

puts in one or two prompts to try and get an answer out the whole way through versus someone building putting in small

32:01

prompts to do little parts along the way. I I think it depends on the stage of your project and maturity. So there is a

32:07

lot of like uh one to like few shot kind of like prompts to do like so if you're doing like analysis or something like

32:14

broader you want to try out like multiple different kind of like modeling techniques or something like that h like

32:19

using the agent here is definitely the the the way to go and that's the way that most people do it. If I wanted to

32:25

do something like make a very specific change to like the colors or like some like extra annotations on a a chart or

32:33

like a specific kind of like limit this uh to only call get like a thousand kind

32:38

of like rows of data instead of 10,000. H that's kind of when people start to use this instead. So it's kind of if you

32:45

have small refinements to an existing block, that's typically when you kind of go for for this mode.

32:52

Now when you when you finish this, if you want to do some production action like scheduling or deployment or

32:58

something like that, uh can you walk us through how that works? Oh, sure. So uh SK scheduling is like a

33:04

super easy. So you have a different kind of like options. So you can go daily, weekly, monthly, custom gives you like

33:10

crunch job if you're kind of a want to. But we could set this up to run if you're a nerd and you can actually

33:17

understand crown. Yeah. Yeah, I know. I know a few of them, but it's like yeah, I'm not a I'm not an expert. Um, so we could set this

33:24

up to run. Uh, we can turn on email notifications uh about it and uh that's

33:29

it. So the version of the uh the notebook at the time of scheduling is now uh running. So that's effectively

33:37

all you need to do to do that. Uh to add a deployment, you're only like a click away. Uh you can add APIs and apps. Uh

33:44

you can do DNS's and there's a cool thing here. So it sets you up with the initial thing, but you could do

33:49

something like uh turn my uh notebook into an app. And h depends

33:57

on how big it is. It might like cook for like 30 seconds or so, but it should come back with like a a full stream app

34:04

uh referencing all of the variables that are interesting. You can set up like multiple pages, everything like that.

34:10

So, it's like um and I guess like tying into some of the earlier stuff we built. So, cloud

34:16

devops, all that stuff built. So, all of that's already wired in here. So, if you have an application, it's in your cloud,

34:22

it's hosted, all of those things, right, Jay? Uh yes. Yeah. And the other thing actually that uh is like new to serve is

34:28

these reusable environments. So you can set up a set of like packages once. So you can have one for like your app

34:35

deployments and then for like model trainings and then when you come here you can just pick the one that matches

34:40

the the type of like app or API that you're uh deploying. The geospatial one is super useful when

34:47

you're mapping and things like that. I used that just the other day. But you own, right? You can create your own. So yeah, you

34:53

just start, you can create new and I can go like new and then you can just add uh whatever packages you want and then you

34:59

can use this across any of your other your other projects. Um so it's like super simple to do like um uh so here's

35:06

our uh all of our code for our our app. So uh what we're able to do then is

35:11

accept it and we're just a click away from a deployment then. Nice.

35:17

And I think the other thing to to kind of talk a little bit about is just um there is kind of a change of process

35:22

when you've got this when when when developing code is this simple or this quick. Like uh we've seen people just um

35:30

doing a level of analysis that is far deeper than they would have if they had to code this stuff manually. Just

35:36

there's no end in the amount of features you could add when trying to do model training or or anything like that. Um,

35:41

not sure Greg or Jay if you want to chat a bit more about, you know, actually changing your own processes because you

35:48

you have these sorts of tools to work with. Oh yeah. I I find that um I do more

35:54

analysis now than I would have maybe in like a year ago because of the agent because it's like um you don't have to

36:00

make such a commitment to be able to see if an idea will work or not. Um so um

36:07

yeah just the the ability to kind of like quickly do it. Uh you can kind of go deeper. Um it'll do a much better job

36:13

on things like visualizations than uh a person would because like typically speaking it's like a lot of effort to to

36:20

write the code to make these things like look look decent. Um, so yeah, I would have find that it's like um uh yeah,

36:28

I've I'm doing more data work now than I was uh anytime over the last couple

36:34

years as a as a result of having the having the agent and trying stuff that I know I don't

36:40

know how to do, but I'm conceptually familiar with. Uh I find that myself doing that a lot. Uh, like my son the

36:46

other day asked me about he was doing some 3D modeling because he he's into 3D printing and he was like, "Dad, I'm

36:52

having trouble getting these edges rounded over in a specific way and I was like, "Yeah, I can I can help with that.

36:57

No problem." So, I took his file and put it in Zerve and said, "Hey, I need to

37:03

round these edges over. Give it a go." And it did. It was remarkable. So just

37:08

knowing about like conceptually like if you know what shap values are or if you know what uh you know uh

37:14

multi-olinearity is or whatever like if you're familiar with the concepts you can type it in and it'll give you the

37:20

code and the implementation for it uh in a way that is you know super useful uh

37:26

and kind of magical. So I find myself doing a lot more stuff

37:31

than Yeah, definitely. And this this is actually just I'll stop sharing in a second, but this is just one of the

37:37

gallery. So we have a gallery the people can publish community and this is so I'm like I like sports. Um and this is like

37:44

a So this is cricket. So this is the IP. Oh, you got to switch. We got to switch. One sec. So share. Can you see it now?

37:51

So it's just Oh yeah, there it is. Yeah. This is so wow. Uh so this isn't this is

37:56

one of like the Zerve community galleries or from the gallery but it's a like a

38:02

deep analysis of like uh all of the different cricketers for picking their fancy uh fantasy league players. Um so

38:10

it's kind of done like a very deep analysis of like uh all of the different like uh uh players. So I just went in

38:16

and um so I have UX because it's a fantasy cricket you Europeans. Oh yeah.

38:25

Well, I think it goes beyond Europe. Yeah. Yeah. Asian. Yeah.

38:30

Yeah. So then it kind of you can then go in and ask the agent. So this is like a big project. You can get the block level

38:36

descriptions, but you can just say like what's happening here and then like what are the top players and like the agent is able to like get all the context and

38:43

give you a quick uh a quick answer. But it was a cool um a cool project to see on the on the community.

38:50

Now we bet we better take a few questions, I'd say. Um there seems to be a number that have came in.

38:56

Yeah. So, uh Rocky Hickey23 says, "How is this different from Jupiter or Google Collab?" Uh I'll take a swing at it and

39:04

then maybe you guys give a go. Uh Google Collab is basically a fork of Jupiter except it's online and it's infinitely

39:10

more frustrating to use uh because it's it's very slow and despite the the name

39:17

including the word collab uh collaboration actually becomes more difficult than Google Collab when you

39:23

use it. Uh Jupiter is primarily we we talked about notebooks. Uh notebooks

39:29

when you say notebooks it's pretty synonymous with Jupiter in a lot of cases but Jupyter notebooks run uh

39:35

locally. They're in memory tools. They uh don't typically have an integration with with an agent or an AI. Uh but they

39:43

are kind of the tool that everybody learns on because they were designed to be classroom scratch pads. And so

39:48

they're very familiar to uh professors to students who are uh writing code and

39:54

then they just sort of take those into the workplace. This takes that to the the next level to an industrial level so

40:01

that uh instead of being in-memory and in being instead of being uh you know

40:07

non-colaborative and local instead of all of that you've got a cloud-based uh

40:12

rockolid bulletproof uh environment that integrates all of the latest LLM and AI

40:18

technology uh and gives you all of these additional capabilities that you could never get out of a Jupyter notebook. So,

40:24

I can pick my compute. I can use GPUs. I can run in the cloud. I can share collaboratively. Uh, we didn't have time

40:31

to show it, but Jason could have invited Philly and I into his canvas and he would have been able to see our mouses

40:37

moving around and we would have been able to all edit and run code simultaneously. Uh, because it's it's very much like a

40:43

Google Docs experience. Uh, when you're when you're working collaboratively in in Zerve, lots of people all at the same

40:50

time. So uh I think the main difference is you Jupiter is is a scratchpad that's mainly

40:57

uh mainly meant for the classroom but has been used in many other places due

41:02

to a lack of other tooling. Uh and Zerve is kind of the evolution of that into an

41:07

industrial-grade development environment for data science. Yep. Nothing to add. I think you got it,

41:14

Greg. Uh okay. Um let's see. Uh but I like

41:21

Uh, let's see. Dumb guy code says, "What does your AI does?" So, I think we're being trolled a little

41:28

bit there, which is fine. Uh, our AI does what AIs do.

41:35

Um, yeah, but I guess in in particular, uh, we we've tuned it for data science, tuned it for the context required to do

41:43

what we refer to as data work. Um, just I'm taking it genuinely, Greg, in case we're not being trolled. I'm I'm taking

41:49

the taking the taking the question. Um, but yeah, ours is specifically for for

41:54

helping you uh get to insights really quick, to do data work, um, and to do what what Jason just did. Yeah. Um, I

42:02

I'm thinking I'm thinking of of uh Simpson's quote about the 500 don'ts and

42:08

10 dos of nice knife safety. Don't do what Donnie don't does. Um, but yeah,

42:13

uh, very very interesting. There's no podcast or live stream that's complete without a Simpsons reference.

42:20

No. All right. Sheetal Andy22 says, "How do I turn something I build here into a real production workflow?"

42:28

Uh, Jason, you want to swing at it? Uh, yeah. So, there's um the few kind of

42:33

methods at the end is one is the scheduling of like a workflow is one. So, you get full version control. you

42:38

get uh kind of like uh all of the the normal kind of like resilience and fallbacks and like retries kind of built

42:45

in automatically and then you're only like uh one click away from an API or a

42:52

app deployment. Um so you can take anything that you've built inside of your uh your notebook and you can

43:00

reference that directly inside of your deployments and then once you click deploy it's uh fully scalable in the

43:06

cloud on your own DNS. So it has its own like URL. Um, and yeah, that's it. So

43:12

you just uh write some code, reference what you've uh uh built in the the notebook and click deploy.

43:19

What about sensitive data? Jason, you want to talk a bit about how how Zerve handles?

43:24

Sure. The data is sensitive. Uh so there's a there's a well the first thing is that it's like fully

43:29

self-hostable. So you can like uh run Zerve in your own cloud. So uh you can

43:35

connect uh it's like a six minute install. So you just click uh the quick start. So if you sign up for reserve

43:40

you'll see a little self-host button you can uh click on and uh then it'll uh uh

43:46

just install it's inside your own AWS. You can run it on Kubernetes. You can run it on prem uh if you want it and all

43:53

of the data all of the workloads everything you deploy sits inside of your infrastructure. Um the other thing

44:00

then is that there's uh data set connections um that you can kind of have once

44:06

there's secret managements that are all baked in. So they're fully encrypted using like vaults. Um and um one thing

44:14

as well I think that Philly was going to touch on around the deployments is all of your code can be version controlled

44:20

through uh git. So any of the major git providers even like GitHub enterprise

44:25

that's hosted like uh inside your own environment you can push all the code that you write from Zerve into into

44:31

that. Got it. So we're not playing around here. We mean b business all the serious

44:38

things. All all the serious things 100%. So yeah, I think from the start um even

44:44

even in our recruitment um like our head of engineering, I recruited the guy that used to say no to tools in our in our in

44:51

where I used to work and it's from the start we wanted the the strictest um controls internally and to be set up

44:58

like we've set up and Jason maybe you can even talk to that a little bit. um you know whenever you kind of self-host

45:04

into into if you self-host in your own cloud um the different control plane and

45:10

the data plane um how that's actually set up similar to the way we use our services and interact with other

45:16

people's environments um so there's the

45:22

which is basically the set of APIs for like uh managing the user access the

45:28

like um the projects the access controls. Uh they're hosted in our uh

45:34

one of our accounts and then uh if you sign up and just want to try reserve uh

45:39

we'll say um you can use our SAS offerings. So this is like um uh will

45:45

like provide the compute and everything but that is set up exactly like somebody that's uh self-hosting themselves is.

45:52

So, it's kind of like the whole thing is once you self host, you get the exact same experience uh as uh we do for all

46:00

of our own users. So, there's no um uh downsides. It's like tested heavily. So,

46:06

there's thousands thousands of users um that kind of go through this uh setup and it's kind of like everything is uh

46:13

optimized for that kind of experience. Yeah. So, it took longer um but it was

46:19

done on purpose, you know, in terms of building it. that was harder to build that way. It would have been quicker to just build a SAS product and then think

46:25

about how you have it work inside an environment. But with like with people like NASA or like financial services

46:32

organizations, we just knew when we got to that it wouldn't work. And so all

46:37

conscious uh decisions uh for the seriousness, Greg, uh to make it work um in the long term.

46:45

All right. Well, uh we've uh we've had a great conversation here. I want to wrap up with a bit of a curveball. Uh you

46:52

guys hear that noise? That's crazy. I want to wrap up with a bit of a curveball here. The large language models are uh kind of changing the way

46:59

the world works. They've certainly changed the way that uh that that people

47:05

kind of live, right? People are now using large language models for everything. Uh what is what's the world

47:10

going to look like in five years? What's going to be the most unexpected or biggest change that you guys see coming

47:16

as a result of the way these large language models are sort of invading every aspect of our lives? Uh, Phil, you

47:24

go first. I'll go first. Well, I think in terms of invading every aspect, uh, I I'm I'm already invaded. So, I was uh early to

47:31

start using it. I um I use I use the various AI tools constantly. So, um I

47:39

had a really strange moment a few weeks ago when uh I use the audible uh uh the audio version of chat GPT when I'm using

47:46

chat GPT and um I have chats with it about you know what I should watch next

47:51

on Netflix and things like that and I was uh having a conversation with

47:57

someone and I was going to reference another conversation I had when uh I was like who was I speaking to about that

48:02

and I wasn't speaking to anyone I was speaking to Chat GPT about what I should watch next. So I think that was the most

48:08

invas in invasive thing that's happened to me in terms of AI. So yeah, there's like the sky is the limit. I I like

48:15

we've really um we've doubled down on the the agents being something that's very different. Um I think we've seen it

48:22

since the uh summer and since kind of launching um into the self-serve agent

48:28

model in like September, just the amount of applications you can have with it. So, I don't know where we'll be in five

48:34

years, but um by no means do I think it's a you know a thing we won't be talking about it. We'll we'll absolutely

48:40

still be on the journey. Bonus follow-up question. Uh have you changed the default voice of chat GPT?

48:47

No, I I got a default voice. I didn't edit it. Still the lady? Yeah, it's the lady. Yeah.

48:53

Excellent. Just Just wondering. All right, Jason, your turn. Um uh I think

49:00

uh probably where they're at at the minute is that they're still um somewhat reactive. So you you instruct and then

49:08

they do. Um there isn't uh much yet in terms of like uh proactiveness and I can

49:13

see that kind of like being something that changes. So kind of like uh something happens all the devices are

49:19

like interconnected and then like it's already taken an action and it just tells you that it's done it more so than

49:24

um I can see even in the the world of like um like a data pip if a pipeline

49:31

fails that it doesn't you don't go and tell it that there's a problem it goes and fixes and tells you it fixed it. um

49:37

thing things like that. Um the interface I think will change as well. Um if you

49:42

trust them more and more, you have to see less and less of the the underlying workings and things like that. So

49:48

there'll be like higher levels and kind of like you start with a kind of like a high level kind of a just like what did

49:55

you do or like what's the outcome and then you can drill down. So I think uh even in coding it'll change it's and is

50:01

changing I think. But Bob Benedict from LinkedIn says, "In five years, demos like this would be done by AI, including

50:08

Q&A, and the audience would be AI representatives of companies that create summaries to suggest next steps." Also,

50:15

nice job today. So, yeah, it's probably true. We are even our Q&A sessions will

50:22

be automated. We should get Papa advisory board. He's got the

50:27

he's got the vision. Yeah. Well, just to wrap us up, I don't know what the next five years will look like, but I can guarantee that in the

50:33

next five years, I'm going to have a Tesla robot in my living room. Uh, I told my wife this and she said, "Well,

50:38

you may not have a wife in your living room if you if you get one of the one of the Optimus robots." And I'm still not

50:45

sure which one I want more. So,

50:50

but all right, in closing, I just want to invite everybody that's listening, uh, go on to app.serve.ai, sign up. It's

50:56

free to use. Uh, you get a free account. you get lots of credits, plenty to use to get you started. Uh give it a spin.

51:02

It's a it's a great application that will undoubtedly increase your productivity. And uh thank you all for

51:09

joining us and we'll catch you next time. Signing off.

51:15

[Music]

51:21

D. [Music]

51:48

All right. [Music]