Latest / Agentic Conversations (formally mlops.community) / Micro Graph Transformer Powering Small Language Models // Jon Cooke // #208

Transcript

Micro Graph Transformer Powering Small Language Models // Jon Cooke // #208 · Agentic Conversations (formally mlops.community)

0:00
Hold up. Before we get into this next
0:02
episode, I want to tell you about our virtual conference
0:06
that's coming up on February 15th and February 22nd.
0:10
We did it two Thursdays in a row this year because we wanted to
0:15
make sure that the maximum amount of people could come for
0:19
each day since the lineup is just looking absolutely
0:24
incredible. As you know, we do.
0:27
Let me name a few of the guests that we've got coming because it
0:31
is worth talking about. If you got Jason Louie, we've
0:34
got Shreya Shankar, we've got Drove, who is product Applied AI
0:39
at Uber. We've got Cameron Rook Wolf,
0:42
who's got an incredible podcast and he's Director of AI at Rebi
0:45
Engine. We've got Lauren Lochridge who
0:49
is working at Google, also doing some product stuff.
0:53
Oh, why is there so many product people here?
0:55
Funny you should ask that because we've got a whole AI
0:59
product owners track along with an engineering track.
1:03
And then as we like to, we've got some hands on workshops too.
1:07
Let me just tell you some of these other names just for a
1:10
moment, you know, because we've got them coming and it is really
1:14
cool. I haven't named any of the
1:15
keynotes yet either, by the way. Go and check them out on your
1:19
own if you want. Just go to home dot ML OPS dot
1:22
community and you'll see what we've got.
1:24
Tungi, who's the lead researcher on the deep speed project at
1:28
Microsoft. We've got Holden, who is the
1:31
open source engineer at Netflix. We've got Kai, who's leading the
1:37
AI platform at Uber. You may have heard of it.
1:40
It's called Michelangelo. Ohh my gosh.
1:42
We've got Fazan, whose product manager at LinkedIn, Jerry
1:46
Lewis, who created good old Lama Index.
1:50
Well, he's coming. We've got Matt Sharp, friend of
1:52
the pod, Shreya Rajpal, the creator and CEO of Guardrails.
1:58
Oh my gosh, the list goes on. They're 70 plus people.
2:01
That will be with us at this conference, so I hope to see you
2:08
there. And now, let's get into this
2:10
podcast. My name is Joe Cook.
2:12
I'm CTO of Data Reception, the Joint Founder Data section, and
2:17
I I like my almond lattes. Basically homemade almond
2:22
lattes, non milk nutters. Welcome back to the ML OPS
2:26
Community Podcast everyone. Glad to see you again.
2:29
I'm your host, Dimitrios, and today we're talking with John
2:32
Cook about the evolving landscape of data management.
2:37
This episode was as much organisational as it was
2:41
technological. We did get into Spark and all
2:44
the nuances there, but wow, what a great way to break it down.
2:50
Listening to John talk about how to align stakeholder values, how
2:55
to understand from a business side of things the value of data
3:00
and how to look at the data in a way that both the technology
3:05
side and the business side can really squeeze the most out of
3:11
was. It was enlightening to say the
3:14
least. So I appreciate how John goes
3:17
into this idea of data handling and I also really appreciate how
3:23
he talks about data products and what needs to be done on that
3:27
side. As I said, it's organisational
3:30
and it's technical, it's part organisational, it's part
3:33
technical. I don't know what kind of recipe
3:36
I am cooking U over here, but if I were to say the parts of one
3:40
and the other, I would say three parts or organisational into two
3:44
parts technical. Hope you enjoy this conversation
3:46
with John. Let us know in the comments and
3:50
please if you like it, share it with one friend.
3:54
All right, let's get into this. So you're in London right now.
4:03
But yeah, before this you were in New Zealand.
4:06
Tell me about that story. Yeah, so my mum, my mum's a Kiwi
4:09
so I spent, I spent 99. I left the UK with a plan to
4:13
sort of do a bit of a a backpack and and that kind of stuff.
4:16
Going to see the family and spent time in the American as
4:19
the month in America and went to Fiji.
4:22
Spent men, spent six months there but ended up being like 3
4:25
weeks and it was raining season. Of course, being a dumb Brit, I
4:27
didn't understand that process. So first week was lovely.
4:30
Sitting, spear fishing on the beach and then two weeks of
4:32
absolute rain in a bath surrounded by Londoners drinking
4:35
beer. I think I could be back home or
4:37
the weather. Weather, you know.
4:38
So I left there and then went to New Zealand, was there for a
4:40
year, went skiing for a full season and full of whole family
4:44
and stuff, and then went to Australia.
4:47
Tend to be there six months. Ended up saying they're 65
4:49
years, five years, bought a house and all that kind of
4:50
stuff. No, Australia.
4:52
Yeah, we're in Australia. I was living in Sydney the first
4:54
time, that said Sydney, and then came back in 2006 and
4:59
interesting enough I met my wife who's from Melbourne, when I was
5:02
in 2000 and 2009 when I was backpacking around South East
5:05
Asia, Lao and Cambodia and this sort of stuff, and came back,
5:08
then went to live in Holland and then we came back here.
5:10
Then we lived in Melbourne, then came back here, So I've been
5:12
back here since, so you've been around A.
5:14
Little bit. We're well travelled.
5:16
Well, fairly well, you know. Some bet this bet the sun.
5:19
Not, not, not as well as others you know.
5:20
So it's a, you know? That's incredible.
5:24
You've got me a little jealous because I want to do the whole
5:26
Southeast Asia and Australia. And you left out something too
5:30
that you told me about. But maybe you purposefully let
5:33
it out. Left it out.
5:34
I want to put you on the spot right now.
5:36
You almost showed up in The Lord of the Rings.
5:38
Was the deal indeed something that happened?
5:41
Yeah, yes. So when I was there in 99, they
5:43
were filming the 1st, the 1st movies on the Peter Jackson
5:45
movies. And I actually lined up to be a
5:47
A. There were two lines in the in
5:50
in Victoria Manufactory in university, and there were two
5:52
lines there, one for soldiers and one for elves.
5:54
And of course, you know, I was in my 20s, I was younger and
5:57
that kind of stuff and walked up the elves and they went and that
6:00
you're a soldier and I'm like, was that if you're not pretty
6:02
enough. So I lined up to be a soldier.
6:04
They took photograph and all that kind of stuff but I didn't
6:07
didn't even make the cut to be a soldier.
6:08
But I yeah, very nearly in that because they basically then the
6:11
whole country was basically roped into you know because
6:13
there was that many people like you know so yeah very nearly was
6:16
being a soldier in law the rings so.
6:19
Man, that is so cool and what an experience.
6:23
You're just there travelling on your year long trip and boom,
6:27
you're like, well let's see, let's try out for Lord of the
6:30
Rings and let's see what happens.
6:32
Yeah, absolutely. So yeah, it was a it would have
6:34
been, would have been, could have been another career path
6:37
knows it. Well dude, what are you up to
6:40
now? Tell us about what you're
6:42
working on and how you got here. Between all of this travelling,
6:46
you eventually started playing around with data.
6:48
And yeah, what are you doing? Yeah I mean I've been sort of
6:52
playing around days before we probably 30 years all used to I
6:54
was actually do virtual reality back in the 90s, believer not
6:57
for British Aerospace and that's my first job out of uni and then
7:00
sort of come through the whole kind of did a lot of low latency
7:03
messaging in late 90s. So door and architecture in the
7:07
in the early 2000s and then we worked for a consultancy,
7:10
financial services, consultancy, deal of risk and kind of
7:12
data-driven projects because banking is all about all about
7:15
data low latency type stuff, you know, Monte Carlo simulations,
7:18
you know billions of calculations that sort of stuff
7:21
went through the whole Hadoop thing as well.
7:23
So did about 10 different Hadoop projects, and a few actually got
7:26
a few of them to work, which was quite amazing really, you know.
7:30
You celebrated by, Yeah, going to Southeast Asia.
7:32
That's like that, exactly right. And yeah, that's what I end up
7:36
running. A data practise and a
7:37
consultancy, fire service, consultancy, left there, went
7:40
into PwC and started an AI data and AI practise, did a lot more
7:44
AI. They did my first generative AI
7:46
project, but not back in 2017. Ohh with PT1 or what did?
7:52
Prior to that we were we were using transforms at all.
7:54
It was like you know we were using hand cracked taxonomies
7:57
and classic machine learning and naive Bayes type algorithms and
8:01
that sort of stuff. The one for a large bank, the
8:03
last Swiss bank doesn't exist anymore which is basically a
8:06
chat bot and on 1500 documents. So that's very small amounts of
8:10
data. So myself and the and the the
8:13
lead data data scientists there a guy who was absolute genius
8:17
who was the next trader built handcrafted taxonomy and we did
8:21
the machine learning to do the inference between you know show
8:23
me the swap rates for the Southeast Asian Bermuda and
8:26
swaps or whatever it is, you know that type of stuff.
8:28
And yeah and then it guided you through the whole chat box
8:31
experience this little sort rates are just what you mean.
8:34
There's a lot of reinforcement learning manual basic feedback
8:38
that's as you can imagine. But it was yeah it was it was a
8:40
surprise to us using transformer type architecture.
8:43
So and there was another one that was done by the team who
8:46
did credit ratings and spitting out like the reports for credit
8:49
ratings for banks. So Apple's being downgraded by
8:52
ex ones and give you the whole report taking news feeds and
8:55
that sort of stuff and rather than kind of classic well market
9:00
data but also these alternative datasets and that was kind of
9:03
interesting to deliver a report it normally took six weeks and
9:06
30 minutes, you know. So yeah, that was another Gen AI
9:09
one we did then again which was, which was interesting then yeah,
9:12
a lot of a I start the work and they're in classic machine
9:14
learning as well regressions and that sort of stuff as well left
9:18
there and then went to Databricks and there were when
9:21
they were very early on. So I was just 2018.
9:23
This was because presumably you had felt the pain of Hadoop and
9:27
you started playing around with Spark and you were like, well, I
9:30
guess this works a little better.
9:31
Well, yeah, absolutely. Yeah.
9:33
It was interesting because the the last project I did in in in
9:37
in PwC just enough was a for actually a regulator and
9:41
actually built a whole architecture and we put Spark in
9:43
it and there are lots of different use cases and one do
9:45
classic prescriptive analytics all the way through to kind of
9:47
machine learning this kind of stuff and they said right sparks
9:50
the engine you and actually the first use case we found the
9:52
small data was self joins and Spock struggled quite a lot.
9:55
So I had this kind of epiphany back then.
9:57
So actually you need multiple engines for multiple use cases
10:00
but you need to be able to tie it together and ecosystem where
10:02
you can do lots of different analytics.
10:04
You formed a graph based analytics or you'll do a machine
10:06
learning this kind of stuff. And actually interesting enough
10:08
the kind of the next after data breaks my form data section
10:14
we've actually built an architecture and a platform to
10:15
actually do that kind of thing. So and then the whole data
10:18
product movement came in came about and the architecture and
10:21
the and the sort of forces align type stuff.
10:24
So and that's what we're actually now kind of building
10:26
out now that whole kind of vision of actually it's not one
10:29
engine to rule them all back to the Lord of Rings reference but
10:33
using genuine data products to actually be able to accelerate
10:36
these sort of things. If I want a machine learning
10:37
algorithm, I want to use Python or use an alarm.
10:40
I want to use, you know, whatever I want to use, if I'm
10:42
going to use a descriptive metric, that's a classic SQL
10:45
engine. They're different engines,
10:47
different things. But we all need to be able to
10:48
work in this kind of cohesive ecosystem, frictionless, quick,
10:53
cheap, you know, fast to be able to move the business.
10:56
So that's really where that whole kind of vision sort of
10:58
started and and it has proved out quite you know quite well
11:02
with inside that inside that project so.
11:07
Yeah, well, so this is fascinating.
11:09
Tell me more about Data Inception and what you're doing
11:14
there. You gave us a bit of the bread
11:17
crumbs. If you say and that I I would
11:22
love to hear about the work that you're doing right now.
11:25
What's your day-to-day filled like, filled with and.
11:28
And then we can go into some of the pains that you're seeing and
11:32
especially around the data products and how that movement
11:35
has come about, the fun you've been having there.
11:38
Sure, sure. So so obviously I started dating
11:40
section because I I saw like the pain I've been suffering a lots
11:44
of people suffer and you know building data and delivering
11:46
analytics is we all know you know taking take months to
11:49
deliver simple stuff in large organisations.
11:52
So I sort of found it. Data section and Vision try and
11:54
solve that problem and with a mixture of business kind of
11:58
knowledge, product management, tech and data trying to put
12:02
those together to actually do fast accelerated delivery around
12:06
these sorts of things and for over 5 years.
12:09
And so data section is that's that's his core mission.
12:12
We do a lot of lot of work around you know classic sort of
12:14
data architecture technology architectures of building spark
12:17
platforms and data brick platforms and all the data work
12:20
like single customer view and lots of those sort of challenges
12:24
we solved. But also we have built together
12:27
a something called a data project pyramid, which is
12:30
actually a process which try and take some sort of concept of
12:33
agile but for analytics and product management.
12:37
So agile, you you, you develop again by go to do research, you
12:43
work with business, you do wireframes and you iterate
12:45
around quickly, this kind of stuff.
12:47
And it's really hard to do that in data.
12:50
And so you know we built this kind of process where we
12:53
actually we did that with with what we call data product which
12:56
you know which analysis components with visualisation
12:59
we're trying to broaden that, sort of bring that stuff
13:01
together. And so we've been doing a bit of
13:03
that with with various clients and it's quite, it's quite a,
13:06
it's a good process but it's still quite manual.
13:09
So we do a product canvas, we do an ideation session and we look
13:12
at what data is needed and do a full prioritisation.
13:14
So the value comes in then do some wire framing, build the
13:16
analytics, that kind of stuff and work the data teams.
13:18
But still it takes a bit of time obviously as as as you can
13:22
imagine. But it's really the thing the
13:25
thing that drives me and it makes my noodle in this in 2023.
13:28
And I was having this conversation 20 years ago.
13:31
If someone like head of sales says show me my total sales for
13:35
last last year, if I don't have that data and I don't have that
13:38
kind of analytics, why does it take me weeks to months to build
13:41
that? Why can't I just at the click of
13:42
a button and actually do that or I want to do let's turn it into
13:45
a forecast like what's myself going to do the next three
13:48
months, you know again goes throw it over to the data team,
13:51
that analytics team three months later comes back, we've missed
13:53
the opportunity there. So why are we still in this
13:55
position where we're actually still spending a lot of time and
13:58
foundational work and still happens if you have you know we
14:01
work with customers trying to get them over that type of hump.
14:04
We're still in that position and we were this 20 years ago.
14:06
It's like well why isn't dial move that move significantly
14:10
there and obviously we'll talk about Transformers and
14:12
Netherlands coming in. I think that is something that
14:14
is really starting to move the dial quite significantly more
14:17
than mindset shift as well as the technology shift.
14:20
But this to me This is why form data sessions to try and really
14:23
just shrink that gap because we're obviously we're a
14:25
consultancy and I spend all my days typically with business
14:30
people, business and tech but mainly business people or I sit
14:32
in the room and go what what what do you need, what do you
14:34
actually need to do to deliver to value.
14:36
What's going to drive your business.
14:37
I don't have, I don't talk about technology I don't really even
14:40
talk about data till the last bit.
14:41
It's like you know what's the key decision you want to make,
14:44
what's the key outcome you want to get to and I do a lot of that
14:46
with you know heads of sales has a risk too far operating
14:50
officers, you Seos as well and that trying to close the gap
14:54
between that conversation and actually delivering something
14:57
that actually they can use to actually make the decisions.
15:00
That's what I'm trying to do with data section.
15:01
That's all we're trying to do, the data section.
15:04
What are some of the main causes that you've seen things take too
15:11
long? When you're talking about these
15:15
reports or forecasts or anything that has to do with data, Why is
15:20
it that it is so messy? What are some main causes?
15:24
O yeah. I mean, we probably spend a long
15:26
time on this and they'll be well documented.
15:28
But the way things are typically find is the staff, the process
15:32
is really just finding what the business wants.
15:36
You know, fundamentally I get a lot of just give me the
15:37
dashboard of my KPIs. It's like well that's that's
15:39
not, that's useful. What I tend to do is actually
15:42
try and distil them down to the two or three things they need to
15:44
make. They're they better, their
15:46
decision, that kind of stuff. And once you've done that, then
15:49
you need to do what are called data archaeology.
15:50
You're going into systems, you're going into the data
15:52
warehouse, you're trying to discover this kind of stuff and
15:55
it's unpicking all that you're brushing off, you know, lifting
15:58
up the rocks, rushing off the type of stuff.
16:00
And if you haven't got the data in in readily available or the
16:03
data is in a particular form, that's for a particular previous
16:06
use case, which I find that a lot.
16:07
And this is one of the challenges.
16:08
I have a lot of data warehouse data platforms.
16:10
They were built at a time where this is the scope of what we
16:14
want to do. Actually we all do something
16:16
completely different now. Ohh, we have to kind of work out
16:19
why the decision made, unpick it, retransform the data, that
16:22
sort of stuff. And that is one of the biggest
16:24
challenges I see even today. It's like we still have this
16:27
concept. We need to get all the data in
16:30
there before we can do anything with it.
16:31
We can't just like let's just take some other transactional
16:33
system, do some prototype and see if it's good enough, see if
16:35
it's got the right quality. You see it's got the right
16:37
inference points, all that sort of stuff before we actually
16:39
onboard it onto the platform. We now we have to onboard onto
16:42
the platform for the first on the infrastructure and that kind
16:45
of stuff. Wait, can I start you there?
16:47
Because if I'm understanding this correctly, it's almost like
16:51
there have been a lot of decisions made in the past on
16:54
how to architect their system. And there's opinions that have
17:01
been made and certain formats and certain schemas that have
17:05
been taken with the data. And if you want to do something
17:08
new, what takes forever is that you have to backtrack through
17:13
all of those opinions that have been set or the decisions that
17:17
have been made and almost untie the knot each time.
17:21
And So what you're saying is instead of letting the data go
17:26
through, these transformations get dumped in the data warehouse
17:29
or the data lake, whatever it may be, just take it at its
17:32
source and then see where it's at, what the condition of that
17:35
data is and almost duplicate it and let it come through a new
17:39
system and and then let it go through the old system and
17:44
justice to see like, where are we at with this data?
17:46
Can we do something with this? So you're almost like diverting
17:50
or making a copy of it. I'm not sure how you go about
17:53
that, but it's it's like you absolutely divert both of them
17:58
and you make sure that the data is there, that you need, you're
18:01
collecting the right data. You can transform it in the way
18:04
that you want. And just so you have a better
18:07
understanding of what that raw data is like, is that more or
18:10
less how you're explaining it? Yeah, except except you're not.
18:12
It's not like an official carnival going through all the
18:14
infrastructure and stuff. You think about it, it's like a
18:16
sample sandbox or sandpit. There's no that's that's that's
18:19
the thing I I found a lot. It's, you know, especially with
18:23
data scientists, even with BI people and stuff.
18:25
You know where they've they said we want to do something new.
18:27
We've had a question asked of our business and we don't know
18:31
what the answer is. We don't know what the data, the
18:33
right data is to get that. So can we actually bring the
18:35
data and have a look at it first and some kind of, you know,
18:37
sandbox environment, not fully onboarded, not fully curated or
18:41
or fully modelled? Yeah, before we actually work
18:44
out if, A, if it's useful, B if it's got the right quality of
18:47
it, you know, if you've got enough statistically significant
18:50
data points. That's the other thing.
18:51
I've done quite a few machine learning projects where we
18:54
couldn't, they didn't succeed because there wasn't enough data
18:57
points. Which one where you know you've
18:58
got 111 outage for one for around IT risk and there's like
19:03
one outage per per year. It's like well you can't do any
19:06
machine learning on that. But we didn't know that we
19:08
actually had to look at the data but we didn't want to wait for
19:10
it to go to the warehouse before we actually saw that.
19:13
So there is, there is an interesting tension around that
19:15
ability to look at the data to be agile and then also to then
19:19
to be able to overlay other parts of the organisation onto
19:22
it as well. But you need to be able to do
19:23
that quickly, inefficiently because sometimes you you're
19:26
sometimes going backwards and force the system teams a lot.
19:28
You know, they go what do you want was well, I don't know.
19:30
But well tell me what you want and you know and you end up this
19:33
kind of backward and forward to what we're trying to do is
19:36
shortcut that before you say right, that's the actual data we
19:38
need. Here's a simple model, right
19:40
that's now push it through the the operational operationalizing
19:43
piece, make it hard and all that sort of stuff.
19:45
But let's do that first. Yeah.
19:47
And it almost feels like the bigger the scale of the company,
19:51
the more of a headache this can cause just because of all the
19:54
different stakeholders that have their fingers in the mix and all
19:58
of the different infrastructure that it has to go through or all
20:01
of the access that you need. And so there are so many
20:05
different ways that this can be complicated.
20:08
Am I missing any? No, it's I mean it's it's it's
20:11
comes down to really the complexity of the estates really
20:14
IT and the data estate you're right and that's typically
20:16
mirrored by you know Conway's law around the mirror.
20:18
The organisation tending tends to get the the the system to
20:22
state that that you know they deserve or they you know that
20:25
mirrors that. So it tends to be because I find
20:27
small companies or smallish companies with really complex
20:30
data that states because they've grown organically, they've done
20:32
some M&A and they're all different systems and spaghetti
20:35
everywhere. And it's this type of stuff.
20:37
And I can get the understand that you know let's get all the
20:40
data sticking to one place then because sort of see what we want
20:42
to do with it. But actually one of the big
20:44
lessons from Hadoop days was actually if you bring all the
20:47
data in as is, stick it into then try and model it as kind of
20:50
stuff, it's you end up you can end up failing because it's not
20:53
the right formats, it's not aggravated the right level.
20:55
What you want to do is bring it in a piece meal and then overlay
20:58
your business use case on it. That's the other thing from the
21:01
loop date is it's suggested all in first and then model it
21:04
before we know what we want to do with it.
21:05
And I think we're now at the point, especially with the
21:07
market coming with budgets tightening last up, we need to
21:10
be able to move play. We need out of not spending vast
21:13
amounts of money on on data infrastructure.
21:15
It's more around how can we get the analytics good enough and
21:19
quickly enough and cheaply enough to actually be able to do
21:21
that, you know, and that to me is, is, is, is the key kind of
21:25
pivot point we've got we're we're now.
21:27
It's funny that you say that these startups that have also
21:31
had just a mess of a data because I've talked to my
21:35
friends in sales quite a bit and I feel like this is a common
21:40
thing that happens to startups with their the head of sales or
21:47
whatever. You start off and just about
21:49
every startup we'll start with like Ohh Hub spots free and it
21:52
does basically everything we need.
21:54
Let's go with HubSpot. They go with HubSpot and then
21:57
once they get mature enough to where they're like all right,
21:59
ahead of sales comes in, the head of sales is like we can't
22:02
work in HubSpot. We gotta migrate to Salesforce
22:04
and then you migrate to Salesforce and did you actually
22:08
need to do that? I don't know.
22:10
I'm not the one to say, but nine times out of 10, probably not.
22:13
It's just because the whoever came in has more experience with
22:18
that new tool. So they feel more comfortable
22:20
and they get to make the decisions and they own the
22:22
budget and so they go for it. But now you're left with a lot
22:27
of complicated stuff because wait, did that data, does it tie
22:30
back to HubSpot? Does it do we can we get the
22:33
report from Salesforce? How does that look?
22:36
What are we doing here when you're talking about something
22:38
as like forecasting from for a head of sales, if we're going
22:41
back to that example, it's just a mess even thinking about it in
22:45
a startup and I can't imagine a bank that's been around for the
22:48
last 200 years. Not absolutely.
22:52
Absolutely. Actually ironically it's not
22:54
it's the, it's the you know the bank sometimes can be can be
22:57
better than the startup because it depends on how it's like I
23:00
said it's the complexity of the of the estate that company comes
23:03
in and how that's that. You know sometimes especially if
23:06
you go to like a challenger bank it's got like 5 system or 10
23:08
systems. It's you know it's it's not all
23:10
you go to a startup that's got 30 systems you know they mean
23:13
that kind of stuff. So it's really that sort of that
23:15
sort of completely, absolutely right, that migration or that
23:18
kind of that complex estate where you've got multiple
23:21
systems, they're migrating between the two.
23:23
And I guess that's the other thing that I'm seeing this.
23:26
You know if you go to more of the architecture side and
23:29
something the data mesh is trying is trying to solve as
23:31
well as this ability of this problem where you're not
23:35
actually having data one side and transactional, what I call
23:39
the transactional side electrical plane and the
23:41
transactional plate. You got transactional systems
23:43
doing your workflows, all this kind of stuff.
23:45
You've got the analytical plane that actually these are merging.
23:48
If you look at your recommendation, there's a
23:50
classic example that's actually an operational listen, I just
23:53
happen to be analytic workload and I've put in a number of
23:56
architectures where people are delivering components from the
23:59
transactional side and the analytical side together in the
24:01
same place. They're actually interacting as
24:03
one because they're all basically operational systems
24:05
that drive your business. And this is the other thing
24:07
that's really interesting around the kind of the data product
24:09
movement. And this is actually when you
24:11
start, start come down to small components which you can deliver
24:14
from analytics teams or from transactional teams or system
24:17
teams or whatever or you can and you start wrapping those, those
24:21
big monolithic systems you talked about Salesforce versus
24:24
you know the HubSpot type thing. You can actually get some
24:27
agility around there because it takes a kind of a very different
24:29
mindset around to be able to do that and a lot of commitment to
24:35
building that and governing it and managing it going forward.
24:37
But actually if you can do that, you get you get a lot of that
24:40
sort of agility, you know, around that.
24:43
One thing that I feel like I've seen a lot as far as design
24:46
patterns go, and I'd love to hear your opinion on it, is that
24:49
you have a data platform that is created at a company.
24:54
Yeah. And then you have a machine
24:57
learning platform that is created at a company.
24:59
Yeah. Does this view that you're
25:01
talking about, especially what you just said, conflict with
25:05
that or how does that evolve it what you're saying, How does
25:09
that differ from this common organisation slash design
25:15
pattern that we've seen? That's a great, that's a great
25:18
question because it's exactly, exactly the point.
25:20
So I think what they're actually you actually combine the two,
25:24
but not as a platform and ecosystem fundamentally.
25:27
So the, you know, data in my mind doesn't live outside the
25:30
use case or doesn't it tends not to live outside the use case.
25:34
What I'm going to be using for and what it means is basically
25:37
that is you're trying to onboard the data, you're trying to use
25:41
the data but in conjunction with that with what you're using it
25:44
for. Now obviously there are some
25:45
exceptions where like single customer views and stuff like
25:47
that where you have to kind of join it across different use
25:49
cases and stuff. But fundamentally to me the two
25:50
things that are intrinsically linked, I think one of the
25:54
things that I'm seeing or trying to move the dial a little bit
25:58
about, let's not try and model the world up front because
26:02
fundamentally the business is, it's all about the business and
26:04
about the business process and what you're trying to do.
26:06
If you bring the data in and married to the use case and then
26:09
compartmentalise it and deploy it at the same time with the
26:12
analytics and the datasets, that's kind of stuff and this is
26:14
what we call a data product. Now the industry does talks
26:16
about the differently, that's what we call actually in a
26:19
product management client sense. Suddenly your whole world
26:22
changes because actually you're not you're not have these kind
26:25
of big data platforms which are just data only trying to trying
26:28
to Co operate almost like in a vacuum.
26:30
And what tends to happen with data teams is not their fault.
26:33
This is the way the way they're set up they're all sort of this
26:36
order takers. You know the business come and
26:38
say I need data for X1 said why do you need it.
26:39
Don't worry about it. Just give it to me.
26:41
It's like well, we don't really know what it is and that kind of
26:43
stuff. And you see lots of things on
26:44
LinkedIn and other things around value.
26:47
When you're in that mode, you've got no chance of, you know
26:49
control of any value because you're just delivering data you
26:51
have no idea. It's all, it's all pain because
26:54
actually if it works, don't get likes.
26:55
And if he doesn't work you'd be you're hammered, right?
26:58
So for me it's being business driven and we talk about
27:02
Daydream actually being business driven.
27:03
This is actually using the data and the analytics and the use
27:06
case all in one and bring it at the same part of the process.
27:09
So the people who actually build in the analytics are bringing on
27:11
the data. Yes, there's going to be some
27:12
governance and overlays and this type of stuff as well.
27:16
But that's that's really where we want we would need to get to
27:19
And that you know you can argue whether that's business teams or
27:21
central teams, business facing central teams, that's that's a
27:24
different, different discussion. But it's really is people who
27:26
are actually building the use cases need to be bringing on and
27:29
managing the data and it needs to be lifted by tech.
27:32
And this is obviously where Elms and stuff really start to come
27:34
in that come into the fore around this because it's it's
27:37
changing that whole mindset. You know, I can upload the
27:39
documents, start questioning, you know, you ask that two or
27:42
three years ago an organisation, can you build that for me after
27:44
12 months and X £1,000,000, right.
27:46
You know, it's that type of what we build a platform by platform
27:50
in this sort of stuff. But now it's actually because
27:52
business people can actually interact with it with open AI
27:55
and this kind of stuff. You know, the light bulbs gone
27:57
on say, well why can't we do this?
27:59
And to me, that's what we need to move to, you know, and it's
28:02
not just obviously interrogating static data through training and
28:05
stuff. It's also being added create
28:06
forecast and create new analytics and stuff in the same
28:09
mode of operation. And that, to me, is where the
28:11
dial really gets me. Well, how do you see the profile
28:16
of the people that are dealing with this data change?
28:21
Because I feel like right now we have the data engineers and it
28:25
almost feels like we're very much speaking to data engineers
28:29
right now. And they're the ones that, as
28:31
you said, they are in a bit of a, I don't want to say lose,
28:36
lose situation, but someone comes to you and they're like, I
28:38
need this data. And then you're like, well, why
28:40
do you need that? Don't worry about it.
28:42
I just need it because I got to make a XYZ or I got to do this,
28:46
but you don't need to know. It's above your pay grade.
28:48
And then they did that person, probably.
28:51
It sounds like to me like that's a data engineer right there and
28:56
what you're talking about in this idea of, Oh well, maybe if
29:01
we start to phrase it differently or if we start to
29:04
view it differently, not phrase it differently but and really
29:07
rally around a different type of organisational.
29:13
Operating model it is. Yeah.
29:15
Yeah, exactly operating. And have this different
29:19
organisational operating model, you are going to see different
29:26
personas per SE, where a maybe as you're saying it's a business
29:31
person that can talk to an LLM or maybe it is someone that is
29:39
very deep in machine learning that also has that data
29:42
engineering aspect. So what do you, if anything, see
29:47
these different personas changing, or if not at all,
29:51
changing? I think it's exactly right.
29:54
So there are a number of ways that we can we can we can slice
29:56
and dice that. But food for us, the way we've
29:58
adopted as much more of a kind of more product management type
30:02
approach, not project products. We're actually building out
30:05
business facing analytics in a kind of product product
30:08
management mode. And what we tend to find is
30:11
basically if you have small teams which are
30:13
multidisciplinary, which actually can hide around like
30:16
you would do like a normal product, like a software product
30:18
or whatever, you've got in data. For us, a small team with a
30:23
business SME, a product manager, we have some called the product
30:26
specialist, which is kind of a mixture of business data and
30:30
tech that actually can then build out analytics and it's
30:33
more like an analyst that thing. No, I will not recommend your
30:42
podcast as the lead engineer TP Bank.
30:45
I'm definitely not wanting other engineers to have all this
30:49
relevant information and I will not give my name which is Lena.
30:54
So definitely I do not recommend this podcast.
31:03
And also a data engineer and potentially potentially if it's
31:06
going to be a complex like use case some sort of technical
31:08
architect, data architect type person, you know but it's
31:11
basically what you're trying you're honing these teams around
31:13
there or spinning them up to actually solve the particular
31:16
use cases with products. So we talked to all right let's
31:19
build a metric on it goes up to the day engineer to build a
31:22
window function, deploy that and then you'll get maybe UX person
31:24
is part of that team as well to actually skin it and put it in
31:27
front of the business that's that's iterated around that.
31:30
You know if if you ask some questions in the SME would step
31:33
in around that and the idea is you're spinning up these teams,
31:36
these multidisciplinary teams with tech that assists.
31:39
And then coming back to your original point around, yeah, do
31:41
we have data platform knowledge platforms and we have analytics
31:44
and data infrastructure IT systems or actually allow us to
31:48
do that to accelerate that. Then we have a we have a very
31:51
different operating model fundamental.
31:53
We have prot facing teams which are multidisciplinary will have
31:56
a look in the update 20 areas and stuff like that as well to
31:59
actually be able to deliver these products.
32:00
You know, so the pessimists don't change but the way they
32:02
operate does and you might have a platform team like data
32:05
platform team with or analytics technology team a bit like your
32:08
cloud team. It's building the infrastructure
32:11
but they are not delivering the pipelines, they're not
32:12
delivering analytics. You have these kind of business
32:14
facing kind of multidisciplinary teams.
32:16
So that I see and I've we've used this a number of times and
32:19
it works really well. So you put something in the box
32:24
when I asked what kind of stuff you wanted to talk about.
32:28
And I'm not going to lie to you, I don't understand shit.
32:33
Let me just read this to everyone else.
32:35
See how many people out there understand from more than 5
32:39
words at a time Here. It's it's like, what do you want
32:42
to talk about? Specialist deconstructed encoder
32:46
decoder transformations, along with data, product management,
32:49
and tech to vastly accelerate prototyping.
32:53
What the hell does any of that mean, man?
32:55
Break that shit down for me because I feel like we've kind
32:57
of been talking about it, but yeah.
33:00
So we have a, we take a step back, we have a process called
33:03
the product period which allows people to create, yeah analytics
33:06
products with UX in front of the business in a very sort of UX
33:11
type, product management type way.
33:14
But like I said mine it's quite manual.
33:16
So you build a product canvas in a lot of companies to this way
33:18
product canvas, you then define what data it is.
33:21
We have a value framework, this kind of stuff and then you throw
33:23
it over to the data seems to work with the data teams and
33:25
analytics teams to actually prototype those things that can
33:27
that can take weeks to months as we talked about Now the we've
33:34
built some technology to actually build and deploy data
33:37
products in minutes fundamentally, but it's always a
33:41
base this type of stuff. So you basically I want to
33:44
create a metric, create the metric who is is, is the sum is
33:48
is the data set you wanna wanna wanna use on it.
33:50
Here's the doughnut chart you want to put it together click
33:53
click, click back. Now there are stacks which you
33:55
can sort of do this to get to today, but there's nothing
33:58
really around where the if you actually want to do that through
34:00
natural language. So if we you know when I'm
34:03
talking to a business person say OHH operating metric, so I have
34:05
to do click, click, click drag drop and that sort of stuff in
34:08
our onboard data sets, build pipeline that kind of stuff.
34:11
Now what we're developing is something that allows you to do
34:13
that in natural language fundamentally and this is where
34:15
the transformer comes in, the decoded encoder kind of
34:18
architecture. So fundamentally we are joining
34:20
an LLM or a specialist LLM or contour of too much about it
34:23
because it's a little bit harsh, but which allows people to
34:27
actually do that whole drive that product life cycle into the
34:30
design, prototyping and onboarding of pipelines through
34:33
natural language. Fundamentally and the the, the
34:36
core of it is you have to come deconstruct the transformer
34:39
because which is the core of the LLM piece and do some specialist
34:44
stuff when we could use like Lama Tim, we've been looking at
34:46
Lamma 2. We're also looking at ChatGPT
34:48
and stuff like that. They're big kind of bee moth
34:50
typing, all singing, all dancing a bit.
34:52
We actually don't need it. We only need like 10% of that.
34:54
Fundamentally, we want to be able to deploy it quickly and we
34:56
want to be able to do things like you know retrain it as we
34:59
go, continuous training and all that sort of stuff.
35:00
So we're looking at deconstructing the actual
35:02
Transformer architecture and actually glueing it to the parts
35:04
that we need to do it to accelerate that whole kind of
35:08
process of defining prototyping, getting stuff in front of the
35:11
business and then being able to deploy it with one Click to the
35:14
to the ecosystem and all the different they get ecosystem and
35:17
then iterate around it. So agility, speed and low cost
35:20
is what we're trying. But this is so through natural
35:25
language. It's not like people are able to
35:28
spin up infrastructure, or is it?
35:31
Yeah, that we'll be able to, Yeah.
35:33
So basically they'll be able to spin up the data products as a,
35:35
you know, like a metric with a with a used piece, UX will be
35:38
able to forecast, they'll be able to create all the types of
35:41
analytics through natural language and then deploy them
35:44
instantly. So basically you're talking
35:48
about really what the business side is asking of the data teams
35:53
already, but now they can just ask it of this LLM and say, hey,
35:59
I want to know what my forecast is for sales in the next three
36:02
months. Yeah, exactly.
36:04
Right. It's, it's, it's, it's that type
36:05
of process. Yeah, Yeah, obviously they'll
36:07
be, they'll be simple to use case, there'll be more advanced
36:09
use cases which will need some some hand holding and some
36:12
people to do that. But yeah, fundamentally it's
36:14
that it's, it's trying to give the business what they've been
36:17
asking for more. I've they've been asking me for
36:18
20 years fundamental you know and you know I feel I've done
36:23
built solutions which are kind of you know rule based and kind
36:25
of you know and workflow based and stuff which which which
36:28
worked well. But obviously with the
36:30
transformer piece is actually you can get that kind of
36:33
feedback very, very quickly. Dude, so this is fascinating
36:38
because on one hand, I I'm wondering about how you can be
36:44
assured of the reliability of these numbers that are coming
36:49
out. Yeah, How are you making sure
36:51
that OK, it's not just sitting out.
36:53
Nonsense. Yeah, that's a very good
36:57
question. So the the, the, the high level
36:59
answer is that we're not getting the LM to spit out the numbers.
37:03
Fundamentally we're interacting with our own infrastructure to
37:06
be able to do that. So you don't you avoid the
37:09
hallucination problem, you avoid the retraining, massive
37:13
retraining problem, even fine tuning and that kind of stuff.
37:15
We're actually we're using some sort of derivative of a rag
37:19
equivalent to be able to actually spit out the numbers
37:22
and actually do the run. We have basically not a number
37:25
of ensemble models which are kind of all integrated as part
37:28
of the the lamps and the natural language piece and the kind of
37:31
the the Gen AI piece that's not actually I'm calculating the
37:36
numbers or doing that comes stuff.
37:38
Yeah, that makes sense because I think that would scare me a
37:42
little bit if it was. Yeah, indeed.
37:44
We all know how much LM's love math and numbers.
37:48
Yeah, indeed, indeed, On the way you think about, you know LMAO,
37:51
you know through the through the attention mechanism and it's all
37:54
probability based. You know, on Open A I've done a
37:56
lot of work to try and then the new version to actually get to
38:00
do calculations, but it's not calculation engine, you know, a
38:05
simple regression or simple kind of even you know, time series
38:10
forecast, that's gonna struggle cause it's not their fault, it's
38:12
designed to do right. So you need to be able to
38:14
integrate that capability into it, but it's actually outside
38:17
but alongside, you know what? Yeah.
38:19
So in a way. And another design pattern that
38:22
I feel like I've been seeing is almost like there's a proxy that
38:26
will take in the whatever the request is and it will route it
38:31
to the right place and then it's it will get that pipeline going.
38:36
So you have and I I like how you said it.
38:39
If I'm understanding it correctly, it's almost like you
38:43
have a model that will understand what the query is,
38:47
and then it will know what to do with that query.
38:50
And so now we're almost venturing into agent territory,
38:55
I guess. Yeah.
38:56
Is that how you look at it too? Yes.
38:58
So it would definitely would definitely in that in that
39:00
sphere where we're actually asking various components to go
39:04
away and do do things which are outside of the outside of the
39:07
LLM. What we're trying to do is
39:10
basically shortcut. A lot of the one motivation work
39:13
does is trying to replicate you know high friction process at
39:18
the moment that's why they you know they they they tend to be
39:21
periodic. We've got this high price with a
39:23
lot of manual can someone figure out what to do fundamentally.
39:26
So it's you know and they're part of the what agents are
39:29
being deployed to do is actually to figure out the way forward.
39:31
You know it's kind of it's interesting if you look at again
39:34
if you look at open air with a Q learning, we I did a lot of work
39:38
a few years ago using agent based modelling which is
39:42
another, you know it's not, it's different but you do have agents
39:46
which singlet system and one of the approaches was used
39:49
reinforcement learning like learning to actually uncover
39:51
hidden States and the data and stuff like that as well.
39:53
So agents are doing that kind of thing.
39:55
What we're doing is we know what the outcome is going to be.
39:58
So we don't actually have, so we don't have to use the agent part
40:00
to figuring out. We can actually be a bit more,
40:04
much more prescriptive around how we do it.
40:05
So it's a lot faster and it's a lot more, you know, it's a lot
40:09
more optimised around that and we don't need a 70 billion more
40:13
to do that, but we can actually. Use so because you know what the
40:18
end state is going to be. It takes a lot of that guesswork
40:23
out of the agent having to come up with Alright, well, here's
40:27
what I think the best plan of attack will be.
40:29
Yeah, because you gotta have a reward mechanism and you've
40:31
gotta have kind of path parcel that needs to follow it and then
40:34
it has to come here, got to retrace and all that sort of
40:35
stuff. We don't need that piece as yet.
40:38
That's the bit where obviously driving because fundamentally
40:41
it's about working with the business teams to actually find
40:45
that, right. And then part of what we do is
40:47
as in data section is actually help the business process for
40:50
doing that. We have to work with the team to
40:52
actually drive out, that is fundamentally.
40:54
So we ask. What's that end state is?
40:56
What, yeah, exactly, right because a lot of business type
40:58
teams we've we've fine actually on very good actually defining
41:01
that clearly that end state. And so we have a whole framework
41:04
where we actually take people through this and we did this sum
41:07
for a company and a couple of months ago we sat with the CEO,
41:11
the CEO, CTO and the head of product and we worked out what
41:15
the four, three or four most valuable data products,
41:18
analytics products that they wanted to build for their
41:20
customers and we worked through that.
41:21
Yeah. So we can do that.
41:23
We did that today, you know, and had we had the technology, we
41:25
could have actually deployed it probably in a couple of days,
41:28
you know and that to me that's the dial that we need to move,
41:32
right. You know, it's interesting with
41:33
those a LinkedIn poll around what a data-driven means.
41:36
That's what data-driven needs to me, where you can iterate with
41:39
the business people, with the value proper value framework
41:42
driving that whole kind of business change conversation and
41:45
then not spend 12 months, two years to try and implement it.
41:49
Internet. Yeah.
41:51
And I laugh when you say a lot of business people don't
41:54
understand because I just think about my own use cases and our
41:59
buddy Dan, who is the data guy in the MLS community, and I'll
42:03
ask him, You know, it would be nice if I could see like, some
42:06
data around this. And this thing is like, well,
42:10
what data are you looking for man?
42:13
I absolutely. Yeah.
42:14
It's like it's actually not asking asking with data.
42:16
You're asking for an outcome and what's that, What's that?
42:18
What's the outcome. So I just get inside your head
42:20
and say actually what you're asking for and what you actually
42:23
mean are two very different things.
42:25
And that again is what we do at day.
42:27
So that's kind of our core USP. We get in the head of the
42:29
business and then drive that conversation.
42:32
Then because we're analytics and data expert and tech experts, we
42:35
can then deploy that quickly build.
42:37
But we have to build stuff a lot of the time and we're trying to
42:40
shortcut. What you're talking about
42:43
reminds me a tonne of when we had Leanne Fitzpatrick on here
42:47
from the Financial Times and she talked about how for them they
42:52
set up. And it's been like almost a
42:54
year. So excuse me if I butcher this,
42:56
but I will put the original in the description in case anybody
43:01
wants to go and watch or listen to that episode.
43:04
Because it it very much is in this vein of even if someone
43:09
asks for data and even if we can supply that data, what we've
43:14
created as like a middle Ware team between, it's almost like a
43:21
cushion between the data team and the business side or just
43:24
like the data askers. It doesn't even have to be the
43:27
business side. It's like the people who want
43:29
some kind of data versus the people that have to supply the
43:33
data. And I think in the data mesh,
43:35
they call this like consumers versus producers, right.
43:38
And so yeah, what they did is they created this cushion of
43:44
people that are very understanding of data and
43:49
they're almost like educators. So when someone wants something
43:54
from data teams, they don't go to the data team.
43:58
They go to one of these hubs or what I was like, oh, it's kind
44:02
of sounds like they're like a human API because the human will
44:06
go and talk to another human and say, hey, this is what I'm
44:09
looking for. I want this data and that human
44:12
API will say, OK, so if we get you this data, what are you
44:17
going to do with it? And that's like the first thing
44:19
that they say, Well, you know, because I want to know how many
44:22
people unsubscribed from this. And then I can send them like a
44:26
follow up e-mail or whatever. And OK, so if you send follow up
44:30
emails, what are you trying to do with those follow up emails?
44:33
Like what is that? It's exactly what you're saying.
44:35
Like what is the end state that you're trying to get to, and do
44:39
we need that data to get there or is there another way around
44:42
it? Or maybe it is that we need that
44:43
data and more. But the person who understands a
44:49
the data ecosystem, yeah. And then be how difficult or how
44:56
easy something is, or what you can do with data will almost
45:00
like bounce a bunch of questions back at the the person who's
45:03
asking for the data before they come to any conclusion.
45:09
Exactly, exactly. I mean we, we even take a little
45:11
bit further because we talk about framing a lot.
45:13
So part of the process we go through is like we're framing
45:16
and we're framing the business problem, are we asking the right
45:19
business problem, which is an interesting one.
45:22
And then we talk about analytical framing.
45:24
Are we asking it in the right way?
45:26
Are we asking what's the type of analytics we need to do?
45:28
Because I'm going to say give me, you know, we got this
45:31
happened quite a lot. Give me a machine learning
45:33
algorithm to do this, this and this and this is like well, no,
45:35
you just need a metric to do that.
45:36
You know it's like it's and the business framing to be like, you
45:39
know I'm going from classic give me my historical KPI's like
45:43
actually. So at the end of the day what do
45:45
you measure on, I measured on efficiency, right.
45:48
So if we give you the top three things to improve your
45:50
efficiency for that day or the end of the week or whatever your
45:53
measurement period is and it's on this the literary data
45:57
points, is that going to solve your problem?
45:59
Yes. So that's that's the business
46:00
framing. And then the technical framing
46:02
is basically the artistic framing is basically are we
46:04
asking the, you know, we don't need AI for every single use
46:07
case, right. Absolutely, we don't.
46:10
Yeah, well can we deconstructed down from a you know a forecast
46:13
for a couple of metrics beneath. We build you for that just you
46:16
have like the your historical matches, but then we are
46:18
forecast on top. We can and we can deliver those
46:20
quickly and easily and you know that type of stuff can we do it
46:23
that way and it's it's those sort of pointed questions that
46:26
that that we that I use all the time to actually to be able to
46:29
ask it again it sounds, it sounds like a similar sort of
46:31
thing but interesting enough if you develop a a sort of product
46:34
management method mindset you that's what you do that's part
46:37
core part of product management is the value, it's driving the
46:41
business conversation and then you bring in the technical
46:43
aspects that's why we like product management and quite a
46:45
lot in this space and it's you know it works really well.
46:49
How do you feel like so aside from what you just said, what
46:53
are some ways that us as engineers or individuals,
46:57
doesn't matter what our background is, can adopt more of
47:02
that mindset and have more success with that mindset?
47:06
There's there's a. Couple of things.
47:07
There's the first one is what's the empowerment in the
47:10
organisation and you you pretty, you can find out pretty quickly
47:12
how how empowered you are because you know someone comes
47:15
last year senior if it comes nasty if you if you can push
47:18
back and are allowed to push back and get listened to, then
47:20
you've got back. If you can't then there's
47:22
there's there's an organisational challenge there.
47:25
The second bit is bound to ask those questions as I said, so
47:28
fundamentally what are you trying to do with it?
47:30
What can I go from like an order taker to what we call consulting
47:34
the trusted advisor. Someone say they understand
47:37
business and I thought I responded to a post the other
47:40
day and you know I've met a lot with, I've seen quite a lot of
47:43
business leaders. They're really sadly understand
47:46
tech and data that drives what they want to do, right.
47:48
They see it as a key tool around do that but a lot of people
47:51
don't fundamentally. So can you get to the point
47:54
where with your business stakeholders, not your IT, your
47:57
business stakeholders to that point where they are trust you
47:59
to want they understand you understand the business as well
48:02
as they do to be able to ask those sorts of questions and
48:04
they trust your advice and this type of stuff as well.
48:06
So can you do it for me? Those are the two, the two big
48:08
things, you know, if you can't do those two things, then you're
48:11
gonna you're gonna really struggle.
48:13
Yeah, because at. The End of the day, somebody on
48:16
the. Business side, your CEO.
48:18
Doesn't care if you're using Kubernetes.
48:20
No indeed, indeed. And so how?
48:24
You're able to abstract yourself away from all the cool tech that
48:28
we all love and talking about and.
48:31
Getting into the. Weeds on and like learning more
48:34
about on a daily basis but then being able to speak.
48:39
It's not different languages, but speak in two different
48:43
positions, yeah. Do you know?
48:45
Do you know the? Hitchhiker's Guide to the
48:48
Galaxy, of course, so the Babel. Fish, yeah.
48:52
Yes. That's what you gotta do.
48:53
We need to be. We need to be the.
48:55
The the Babel fish where we can actually translate between the
48:57
thoughts. It's not necessarily the voice
48:59
the thoughts between people you know and but also the other
49:03
thing I would say come back to some power.
49:06
The thing I, you know I've lived, I've lost count how many
49:09
analytical solutions I've delivered to the business that
49:12
would change the dial, but they haven't been adopted because of
49:14
all sorts of other reasons. You know we know that called
49:16
politics of that sort of stuff. So that's something you
49:18
definitely need to test. And your organisation, what's
49:20
the appetite for actually being data-driven, that kind of stuff?
49:23
You're how do you test that? But you you basically you.
49:27
Push back and you actually ask those questions.
49:28
You know, like I said, you say that actually let's understand
49:31
what you're trying to do actually if you did it a
49:32
different way or here's an improvement, you know this all
49:36
will drive you know 10% on the on the bottom line on crucial
49:39
conversion rates is kind of stuff that's like actually we're
49:41
not interested in that. But you know, that's a very good
49:44
way of actually testing whether there's an appetite to be
49:46
actually, you know what I say, analytics is a business change
49:50
process or should be if it if you test it, it isn't the
49:54
business change or the scene of that or something else reporting
49:57
or this kind of something we have to do.
49:58
That is a really good litmus test of how successful an
50:02
analytics person is going to be in the organisation.
50:05
Yeah, that is. Some fascinating stuff to think
50:09
about, because you hit on such a clear point where you can do
50:17
everything right and produce, you can create some kind of
50:22
output, but if at the end of the day the adoption isn't there,
50:27
it's worthless. Yeah, exactly right.
50:30
Exactly right. And it's, it's interesting.
50:32
I see lots of like on LinkedIn around and other other posts
50:35
around, you know, you've gotta make sure your data products are
50:38
adopted and valuable. There's kind of stuff like,
50:40
well, if you've built them and you don't know that you've done
50:44
my mind, there's gonna be a challenge, right?
50:46
You build them because there isn't enough because there's a
50:49
need, There's actually a business need to be able to do
50:51
that. And I think that's something
50:52
that in the data world we need to try and have.
50:54
We're still struggling a bit with this kind of again and it's
50:57
back to this kind of data in one place to start modelling it,
50:59
this kind of stuff. It's ohh, the business aren't
51:01
using it. I've lost count of how many
51:03
times I said, well we've built this data platform, we've built
51:06
this data infrastructure, we've done this whole kind of stuff
51:08
and now it's using it. It's like well that's because
51:10
you've gone from a technology focus angle rather than a
51:13
business change because the business angle is actually is
51:15
forget the technology start with this workout what the business
51:18
needs to do and get their buy in from before we do anything you
51:22
know that's fascinating to build it and we will come was was you
51:25
know it was kind of that could have been a big graveyard in the
51:27
Hadoop world. You know that's that was hung
51:29
the big the big promise of Hadoop build it noble but it all
51:31
in one place and all go and it's like no actually we gotta pivot
51:34
to that to actually start talking about business
51:36
conversations and work out what tech and data we need for those
51:38
businesses. There will be some
51:40
infrastructure, there will be some cross some foundational
51:42
stuff we need to build. But it's not like 1216 months or
51:45
18 months because the market might wear it.
51:47
Now with budget shrinking, those can't stuff.
51:49
Yeah, yeah and. And who knows?
51:51
What kind of all of them will be out by then?
51:54
Ted, all of the CEO C-Suite is going to say, well can we just
51:58
do this with AI now And he's like no, but shit yeah there's
52:03
like I can't imagine all of those people that are building
52:07
their data platforms And then when ChatGPT came out everybody
52:11
all of a sudden thought oh you know we could just throw some AI
52:15
at this and how many times the product managers or had to go
52:21
through and figure out story around that and say absolutely
52:25
yes but no. It's also the interesting though
52:27
the 1st. Time in my career and my 30 year
52:29
career, I've been through a number of epochs and on Internet
52:32
and ecommerce and big data and the first time we're actually a
52:36
new world changing technologies come out, we're business people
52:40
can play with it directly, you know, So like they go to chat to
52:44
give me a business plan or let's generate this e-mail.
52:46
But you know I can't really think of any other major
52:49
technology advantage. You know, the advance that's
52:52
happened in the last in my lifetime anywhere where the
52:55
business people and maybe the Internet maybe, you know, but
52:58
there needs to be leader bill stuff.
52:59
It's actually there. I can do all sorts of things.
53:01
I can go and chat TP or anthropological or whatever it
53:04
is and the business before I want that because I can play
53:07
with it and why can't you build that?
53:08
For me it's very front and centre, whereas before it's been
53:11
like oh there's some sort of black magic going on, some black
53:14
box type stuff. We can go and build it and the
53:16
promise happens and it doesn't happen.
53:17
This kind of stuff, you know, But now it's the business
53:20
people. I've got their hands on it.
53:21
Oh my word when asked to take people.
53:22
Well, how do we to fulfil that promise?
53:25
But it does what it also means that the business people are now
53:27
starting to think of all those things they've been asking for
53:30
for years and haven't been under limit for whatever reason.
53:33
It really, you know, it's it's it's the light bulbs.
53:37
Gone on fundamentally around. You know, it's that kind of
53:39
actually, Oh my Yeah, you know, that's that's that's starting to
53:51
permeate now, which is first of all, I've really seen that in
53:54
that kind of holistic way. Yeah, it.
53:57
Is funny. You talk about seeing people on
53:59
LinkedIn, mentioning the adoption and building things
54:04
that and you have to have adoption for your products and
54:07
all that and how you said if you don't know that before you build
54:14
it you failed. If you don't know that there is
54:17
an actual value and use case and people are are jumping over one
54:22
another to get to this product, then your taking on a lot of
54:28
risk. And my buddy Adam said that one
54:32
time when I was in a meeting with him and he was showing me
54:34
what he was tinkering on, I was like dude this is so cool like
54:38
what have you been? How have you built this or what
54:42
have you been doing? Blah blah blah?
54:44
And one thing that stuck with me from the conversation that we
54:47
had was him saying I'm tired of making products that people
54:51
don't want. And it was.
54:54
Just that like so. Now all he does is he works a
54:57
little bit, maybe changes a few things and then goes and talks
55:00
to as many people as he can to see what they think of it.
55:04
And then maybe he'll change a few things and then go back out
55:07
and talk to as many people as he can.
55:09
And that is really the way forward.
55:12
That is super valuable. It just feels like the right way
55:16
of doing it, right? Stuff at Wells run like that,
55:18
right? Soft World is absolutely run
55:21
like that and the big companies, larger companies get them
55:24
walking, the bigger they get the the slower that type of thing
55:26
happens right or doesn't happen at all you know because it's
55:28
it's it's a very different culture.
55:29
So you know we I talk a lot about the UDA loop, you know the
55:33
observe, orientate, decide, act which is obviously from the
55:36
fighter pilot thing. You know obviously the speed of
55:39
that depends on the speed of your business obviously, but
55:41
it's it's that kind of scientific method.
55:43
Let's let's, let's work out what's happening from a data
55:45
perspective, another perspective.
55:47
Let's orientate this data and work out what it means to us,
55:51
you know and then decide then act and do it again and and
55:53
iterate and that the measurement is basically talking to customer
55:55
talk to users, talking that kind of stuff.
55:57
You know if we if you didn't do that and start up well you know
56:00
you'd you'd be dead right. You know you know that startups
56:03
will fail because you don't spend typically.
56:06
I mean it's not always the case, but that's actually spend two or
56:08
three years and millions of pounds developing something then
56:10
and tested customer right. You just don't do it right.
56:12
Yeah, well. John Man, this has been
56:15
fascinating. I really appreciate you coming
56:16
on here, and for everyone that is not already following you on
56:21
LinkedIn, I highly recommend it. You put out some great content
56:24
and I always enjoy when you pop up on my feed.
56:28
That's it. That's where that's the podcast
56:30
for today. Brilliant.
56:31
Well, thanks so much. Do we just have?
56:33
We all been a bit interesting chat and had a lot of fun, so I
56:35
really appreciate it. I'm Emmanuel.
56:42
Mason, Machine Learning engineer at Stripe and author of Building
56:46
Machine Learning Powered Applications.
56:48
And if you don't want your machine learning models to
56:50
explode, well, you should subscribe to this.
56:53
Podcast.