Latest / Agentic Conversations (formally mlops.community) / Leadership on AI

Transcript

Leadership on AI · Agentic Conversations (formally mlops.community)

0:00
And you are not just providing A functionality or feature to the
0:03
business anymore, you also need to help the business to go
0:06
through this big transformation. When we share these tools to our
0:10
colleagues, they came back to things that we didn't expect.
0:13
Yeah, well. Practically, the problem didn't
0:15
start with AI, we had that problem in the past as well, so.
0:18
You mentioned that it's hard to measure.
0:20
It's not hard to measure. The question is, are you
0:23
measuring the right thing? We can start out with just the
0:32
idea of how you are thinking through enabling your different
0:36
departments and the wider companies to be more productive
0:40
with AII. Mean the role has changed quite
0:43
a lot for me as well because up until 2-3 years ago, your focus
0:48
as a CTO is to basically provide the functional to the company.
0:52
Needs comes from product, comes from marketing, comes from any
0:57
department and also you are responsible for the scalability,
0:59
stability of the platform. You have to make people trust
1:04
what you are building in. In the last two years with AI,
1:10
with Jet GBT being available to everyone, you also realise that
1:15
oh OK, this, this is big and you are not just providing A
1:19
functionality or feature to the business anymore.
1:21
You also need to help the business to go through this big
1:24
transformation about how you're thinking, how your processes,
1:29
how your way of working needs to change.
1:31
And as as you are the responsible person for the
1:35
technology in the company overall, you have to basically
1:39
leave your comfort zone and then push the boundaries across all
1:44
these departments to help them understand what is coming up,
1:48
how big that will be for those departments, for the business,
1:51
for the ecosystem and for the for the world's.
1:55
And also in the background build the foundations and the
1:58
platforms so people can operate on these things as well.
2:01
So. You have to become a shepherd.
2:03
Yeah, exactly. Exactly.
2:04
So you're not responding to requests anymore.
2:07
You start to push the boundaries of the business more than you
2:10
have ever done. So it was a change.
2:13
So a couple of additions, let's say, because I can see
2:16
everything that that Merit has said so far.
2:19
The few decisions that we made even before Charge PT is that
2:24
all this technology is large language models, independent of
2:27
the fact they're not perfect, independent of the fact that
2:29
this is, you know, it's going to take years before we get
2:32
something which is foolproof and so on.
2:34
But they're so new and so impactful that you need to find
2:39
a way to learn them. And then you can decide to
2:42
mandate to create education and so on.
2:45
Or you can take the opposite direction, which is let's give
2:49
everybody the tool so they can experiment bottom up.
2:53
And I totally convinced that for every organization that I see,
2:56
this is one of the best options you can possibly have.
2:59
It's provide within boundaries that the boundaries being the
3:03
guardrail, such a way that you do not make catastrophic
3:06
mistakes and so on, but give the tools that people can decide how
3:12
to create agents for themselves, how to in fact create the things
3:17
that they need for their work, right?
3:19
And the reason why we did that is because long ago GPT three, I
3:24
think it was really, really early stages when we share these
3:28
tools to our colleagues, they came back to things that we
3:31
didn't expect use cases that they would.
3:33
I can't invent that thing because I'm not in that line on
3:36
business. I will never be able to figure
3:38
that out. So by giving everybody, it means
3:42
literally everybody, tools that can experiment.
3:45
There are two things that we bank on.
3:47
The first one, they will create something useful for them.
3:50
So there's going to be a productivity quality of work
3:54
that increases, which then if you multiply for the entire
3:57
organization is going to be material, very material.
4:01
And the second thing is that there is a distinct advantage in
4:04
having everybody understand where these tools work, where
4:08
they fail, how I can tweak them to make them work where I need,
4:12
because then they can apply the same intuition to the
4:15
applications that they developed for our consumers, right?
4:18
So these two things tend to become, let's say they are the
4:23
same thing in the end, but they're very important.
4:25
So collective discovery on the one hand, but also collective
4:28
awareness on the other hand. Actually, I think one thing that
4:32
I was shocked by was when Paul was telling me how you guys
4:36
encourage most developers to use various types of coding agent
4:42
tools. Yeah.
4:43
And it's for the sole fact of, hey, if there's something
4:46
valuable here, you should know about it and you should choose.
4:48
We don't want to say, oh, we're going to mandate you can only
4:51
use cursor, or we're going to mandate you can only use XYZ.
4:54
Insert your favorite coding tool, cloud code.
4:57
But we want you to be able to find the one you like and use it
5:02
if it gives you that productivity.
5:04
Yeah. So I'm, I'm, you should weigh in
5:06
on this one as well. But there are advantages in both
5:09
sides. On the one hand, there is a big
5:11
advantage of one tool. Everybody standardized on that
5:13
one. You can get better conditions,
5:15
commercial conditions, and also a lot of knowledge that can be
5:18
easily shared, right? So because you just create,
5:21
let's say, a lot of knowledge there.
5:22
However, the reality is that these tools are still developing
5:27
at this moment. There is a winner, but next
5:29
year's might be another winner. So there's going to be, let's
5:31
say, an evolving landscape. And you also want to make sure
5:34
that you continuously test and stretch all the tools out there
5:38
for various reasons, because for many use cases, you better use
5:42
different tools. It's just a reality of that.
5:44
But second, because it is evolving.
5:47
So you don't really bet on anything, but you bet on a way
5:50
of learning these tools, right? So in that case, having a
5:52
certain degree or total degree of freedom is fine.
5:55
The other thing that we are not really, let's say, compromising
5:59
on is to to push everybody to use these tools, right?
6:04
So there's a deliberate mandate to experiment and get to 100%
6:09
adoption as fast as possible, for the same reasons that they
6:12
discussed before. Yeah.
6:15
So we didn't exactly follow the same approach in terms of just
6:18
just let everyone use everything.
6:21
We started with GitHub Co pilots.
6:23
We also launched Amazon Q developers.
6:26
We are right now launching cursor.
6:31
I think the Android engines are using Gemini and I know there
6:35
are lots of other tools as well available.
6:38
The reason why we didn't just let's all is to also get provide
6:44
some focus to the organization. And also I share some of the
6:46
learnings about how to make the best about a certain tool
6:50
instead of jumping from 2:00 to 2:00, because there's, there's
6:54
also a trust element on, on if an engineer will trust to
6:57
technology. And for building the trust, I
7:02
think you need to spend some time like you need to give the,
7:04
the credit to the, to the, to the AI or to whatever it is
7:07
that, well, it will not solve everything at first go.
7:11
So you need to get better with your prompts.
7:13
The AI will get better with the models.
7:16
And there are things that it will be great on doing and it
7:19
will, there will be things which they're really not good at.
7:22
So, and if our, our assumption was that if we let them just
7:27
just jump from one tool to the others, we may lose that that
7:30
momentum on investing right time.
7:33
And also we would like to provide some guidance about how
7:37
to get best about a certain tool, internal trainings, having
7:40
the providers coming and, and telling people how to how to use
7:44
those tools and commercial side. It also helps with the
7:46
negotiations as well. If you if you, if you go with,
7:49
with a few of them. But I, I fully agree, like the
7:52
idea is not not to limit or, or make it very strict so that
7:57
people feel like, OK, I mean, this company's just just not
8:01
evolving. This company is against these
8:03
two. So if you give that feeling and
8:05
if you just push for a certain thing, I think engineers are
8:08
very smart to understand that this company is not going to be
8:11
in a way to this company will disappear.
8:12
So, but also giving some guidance helped, helped believe
8:16
to us, we were quite pushy about the adoption of these tools as
8:23
well. So in the all hands, as I was
8:25
very clear that like you guys have to use this.
8:27
So there's just no, there's no option because it's not about AI
8:30
replacing you, but someone using AI will replace you.
8:33
So it's not about you losing a job on just it, you are losing
8:36
your profession if you're not using these, these things
8:38
because these things will not disappear.
8:40
These things will be part of our work.
8:43
And, and the adoption has picked up very much like I think we
8:46
have more than 95% of the engineers using these tools
8:49
every day for their day-to-day work always If we didn't, we six
8:54
months ago we made a calculation that we believe 3040% of the
8:58
code generated is AI generated on, on the production systems.
9:03
And it's very difficult to measure that because you have
9:06
suggestions they can copy paste, they can use other tools as
9:09
well. So you cannot easily 100%
9:13
accurate with this number. But the, the reality is we also
9:16
didn't see a 30% velocity increase in, in the organization
9:19
because software development is just not, not just coding, it
9:24
has lots of functionalities as well.
9:26
And all these things are getting better over time.
9:29
But the trust, the learning, the, the way we are operating is
9:33
changing over time. And I think I mean we are in a
9:37
very strange and interesting and exciting time.
9:40
You bring up a really good point around 30% or 40% of the code is
9:47
generated by AI. It's hard to track the metrics
9:50
on that. I want to get to that in a
9:52
second, but I also want to raise the other point that you said
9:55
it, which is we didn't see a 30% increase in velocity because
10:01
that's not the bottleneck. And I have a joke.
10:04
I'm always saying that the bottleneck usually is the
10:06
aligning part. It's not the actual like hands
10:09
on the keyboard part. The aligning takes much longer.
10:12
Well, it's getting better as well as as you said, there's
10:15
just so many factors in in place on affecting your total outputs
10:20
capacity or velocity. That depends on on the coding,
10:24
that depends on the alignment with your team members, with
10:26
other teams, with product manager, waiting for UX, waiting
10:29
for QA. So but I, I think AI is helping
10:33
in all areas in parallel. So it will be a mistake to just
10:38
bet on certain process and just optimise it to a maximum and
10:42
then ignore it. I think it will all evolve at at
10:46
a certain pace. But also the reality is behind
10:50
the scenes there are also new work that is coming up about
10:53
being more AI ready or AI available as an organization, as
10:59
a as a as a technology as well. So the way we have architected
11:02
these platforms were based on the fact that's always the
11:06
software engineers will will scale themselves, the teams will
11:09
scale themselves. We started with monoliths than
11:11
microservices. And probably right now we will
11:14
see a different shifts of half software will be architect so
11:18
that AI can generate code much, much more satisfactory than the
11:24
the than the limitations of today's architecture.
11:27
So that means additional work will come to the product and
11:30
engineers as well. So you mentioned that it's hard
11:32
to measure, it's not hard to measure, it is.
11:36
So there are many tools to measure productivity of coding,
11:38
right? So as to then and, and there are
11:41
companies that have tools for, for that job.
11:44
So you can do that. The question is, are you
11:47
measuring the right things, right?
11:48
So the fact that you have, let's say, more PRS, let's say, or
11:52
things like this is that, let's say good measurement of
11:54
productivity. So that remains the The thing is
11:59
that in any case, coding is just one of the work streams of a
12:05
company. So if you look at the entire
12:06
company, there's engineering that does many things, including
12:10
coding. So, so there's a group of people
12:12
that are contributors to code contributors sort of thing.
12:14
But then this type of not automation, but in any case
12:19
agentic use to improve your work can be applied everywhere across
12:22
the organization. Then interesting part of coding,
12:25
but that's similar to what you see in legal profession, what
12:29
you see in customer support, you see in some ways also in
12:32
marketing is that software development is a well understood
12:37
discipline. So there are schools, there are
12:39
trainings, they are coding tools, of course, you know,
12:44
they're best practices and so so people more or less know how to
12:47
do it and then others can judge if they're doing well or not and
12:51
so on. So there are parts of the
12:52
organization where these kind of agentic augmentation follows a
12:59
fairly well defined path just because the discipline is well
13:02
defined. And then you have a long tail,
13:05
an extremely long tail of how to call it is sort of the way to
13:09
the way I look at them is tiny total addressable markets while
13:15
coding it is a large total addressable market is within an
13:18
organization you have thousands of tiny total addressable
13:22
markets where the total addressable market might be 1 or
13:25
10 people, right, almost in the individual basis.
13:29
And they can use these tools to make their work more impactful,
13:33
can be faster, it can be better, it can be whatever metrics you
13:36
have. Now that's another very, to me,
13:39
it's a very interesting segment of a genetic automation, which
13:44
usually is not captured by let's say one or two companies, but it
13:48
is created by every single worker bottom up.
13:52
So that's one of the reasons why we have been working so much and
13:57
pushing so hard in getting in the hands of everybody in the
14:02
group tools such a way that people can create their own
14:06
agents. Because there might be 3 people
14:08
in the marketing department that struggle with that problem.
14:11
They repeat the same thing over and over and over, and now they
14:13
can actually do it much faster so we can do something else.
14:17
How do I get that problem if I'm not there, right?
14:20
So there's no way that they can invent.
14:22
It's not like coding. We know what good looks like,
14:25
but here I don't know what good looks like.
14:27
They know, right? And they are the only ones that
14:30
they can put together something that makes their work better.
14:33
So it's a tiny addressable market, but they have it 1000
14:37
times, right? So if you have a very low
14:40
barrier to creating these agents, then you can create
14:42
thousands of automations across the organisations, which in
14:47
total, they're really impactful, right?
14:49
So that's the long tail of, let's say, agents in, in, in, at
14:53
work. And it doesn't make sense for a
14:55
company to create a product to service the three people in
14:59
marketing, of course, because. There's no market.
15:01
Exactly. There's no market, so you need
15:03
to have a tool that makes it extremely simple and extremely
15:06
easy to create these automations for you.
15:08
Then it makes economic and practical sense, right?
15:12
And that's the reason why it was not done until recently.
15:15
It's probably a good segue to go into your bold vision for 30,000
15:21
agents in production. What is it next year?
15:24
March. March.
15:26
Yes. Where Where you are right now?
15:29
We're 18. 18 OK, yeah, that's. We're getting there much faster
15:33
than that. So yeah, we're going to go
15:34
there. There was a graph.
15:36
There's a graph. It's exponential, so it's.
15:41
Let's go back and and, and, and think about why is that a good
15:45
idea. Even if it looks silly, I think
15:49
it's a very good idea because it has to do with change
15:54
management. It has to do much less with the
15:57
agents, but it has to do much more ability of everybody of
16:00
removing the barrier that they believe for agents.
16:03
So back in July when we we have seen the agents creating with
16:07
Tocan, the platform that we use for that sort of plateau, in
16:11
spite of the fact that had all the functionalities they connect
16:14
to MCPS, they could connect with let's say dozens of internal
16:18
systems, all the people were not adopting more than the technical
16:22
people, right? So only technical people were
16:24
sorted is and then we had to figure out the reason.
16:27
The reason was extremely simple is because everybody else
16:31
thought that this is a tool for engineers.
16:33
Engineers can create agents. So we have to wait for them.
16:36
So it's going to come in and help and so on and so on and so
16:38
on. So we had to remove the barrier.
16:40
And how do you remove it? By sitting with the marketing
16:43
people, by sitting with HR people and creating these agents
16:46
in 30 minutes, in 15 minutes, not in a week, Right.
16:50
And once you go through the process, then you see that you
16:54
just say, wow, is that's it? Is that all it takes to create
16:59
an automation makes a difference?
17:00
Yeah, that's all it takes, right?
17:02
It is at this point in time is simple.
17:05
So that what made a difference and they started going up.
17:09
A few other things happen when you're at that point that if you
17:13
have created an automation that helps you, you're going to tell
17:16
somebody else, hey, I've done this one and is going to help,
17:19
why don't you do it too, right? And so on and so on and so on.
17:22
So that mechanism goes into, let's say, a percolation of
17:26
knowledge that enables everybody to create the age what we don't
17:30
know how useful and how good these agents are, right?
17:33
We are going to focus on that afterwards.
17:36
First, we make sure that everybody feels a tease that
17:39
within very specific boundaries, right?
17:42
So you can't do what you want. You have very strong security
17:45
provisions, but within those security provisions, everybody's
17:49
encouraged to experiment as much as they can in such a way they
17:54
can become the owners of those tools, right?
17:57
So there's not an engineering tool, it's a tool for everybody.
18:00
Once you get there, you start getting into what are the good
18:04
agents, how can make them better, right?
18:07
How can you change the processes then that generated this
18:11
inefficient in the 1st place. So how can you change the
18:13
organization? It's very hard to get there if
18:16
you don't have everybody on board, right?
18:19
So 30,000 agents, in fact, means it's a change management
18:23
process. It's a cultural process.
18:25
It's a, it's a, it's an organization thing.
18:28
It's nothing to do with the technology itself.
18:30
I, I actually really, really like that approach and I can, I
18:33
can just add to what you said, although we just became part of
18:36
process. So Tokan is not available to our
18:39
organization yet, but we will help the 30,000 for sure.
18:42
So you will make it before. March, hopefully, but to get
18:45
over that. I I like the fact that there are
18:47
two things. One is you put a bolt idea there
18:52
which no one can ignore and that creates a question mark in
18:56
people. Like what is that?
18:58
Why 30,000? What is talk on like?
18:59
Should I be worried about that? Should I get a training?
19:02
So you create a demand, you create an expectation.
19:05
You can create a question mark in in people's.
19:08
If you don't go bold, if you don't make it challenging, it's
19:11
very easy for people to just move on, ignore, and focus on
19:14
their own work. And the second one I also agree.
19:17
Like I have a daughter, 8 years old.
19:19
She's she's learning to read and write.
19:23
And my KPI is how many pages she reads a day.
19:26
I actually don't care if she's reading the stories that I want
19:29
her to read or not. But at this point, I want her to
19:32
read, I want her to understand, I want her to get better, more
19:36
comfortable with what reading is.
19:39
And then when she gets more comfortable, then we will
19:42
discuss about which stories, which values she should she
19:45
should get about it. But right now, similar to
19:47
probably what her was doing is you need to create the first
19:52
level of understanding about what is going on.
19:54
And everyone needs to jump into that boat by forcing themselves
19:59
and the value, how you are going to get the maximum value out of
20:02
that is the next stage. So I think that's that's that's
20:04
the right thing. Capture the imagination in a
20:06
way. Yeah, absolutely, absolutely.
20:08
And then as I said before, we always go back to some let's say
20:12
of the tests that we did in, back in 21.
20:15
It was really, really early on with, with LLMS and and people
20:21
were really coming back, right. They didn't, they haven't seen
20:24
these tools before. So that we stitched together
20:26
very simple user interface on top of GPT 3, I think something
20:31
like that. And we gave it to our colleagues
20:34
and do something because it looks interesting.
20:36
We never been able to do good stuff with with NLP at this
20:39
level. So this is really bad at this
20:41
moment, this the time tools really bad, right?
20:44
But they were really promising too and they were improving
20:47
really fast. So if that trajectory holds,
20:50
then you're going to have in one or two years something which is
20:53
really surprising and we never had before.
20:56
But how do you use it? Do I know?
20:58
So we had the ideas of course with, you know, we repeated all
21:01
the NLP use cases that we had before with, with LLMS, but then
21:06
we gave it to about 100 colleagues at that time, sort of
21:09
some really rudimentary GGBT where you could test and so on.
21:14
And people were coming back with things like they created exams
21:18
for schools. They're completely personalized.
21:21
They put, let's say, a small CV of let's say students in the
21:25
prompts of the doing experiments like that and so on.
21:28
And and how did you think about this country?
21:30
It's actually a great idea. It's a great idea.
21:33
Did I have that idea? No.
21:35
And so many other things. So that what convinced us that
21:38
because of the novelty of these tools, because we never had
21:42
tools similar like this one, you have to get, you have to
21:47
experiment bottom up if you want to find out, let's say, what you
21:51
can possibly do with these things and but also all the
21:54
errors that you can find. This mechanism of, let's say
21:58
collective experimentation is one of the greatest sources of
22:02
threats, potential errors. So that's how we learn where we
22:06
should not go, right? Because then you don't have one
22:10
team figuring out. In fact, you got 30,000 people
22:12
doing red teaming all the time, which is also very useful.
22:17
Huh. Well, yeah, because you're
22:18
scratching your own itch. At the end of the day, I have,
22:22
whether it's professional or personal, I have this thing that
22:25
I want to do. Let's try and see if I can build
22:27
an agent to do that. Absolutely.
22:30
And then and then you can see that somebody else has done an
22:33
agent on education just to tell you that example, but you can do
22:37
the same thing for customer support.
22:39
You connect the dots, right? And then that dot connection is
22:43
what you need. Because we just don't have an
22:45
intuition yet about, let's say, the full potential of this tool.
22:49
So how do you get to explore the boundaries?
22:51
Just stretching them as far as you can, yeah.
22:54
It feels like there is a lot of the change management piece here
22:58
and being able to encourage the employees to spend time with it.
23:05
As an employee I would probably think I'm so busy with other
23:12
stuff to put the time in, even though I know that later on
23:19
maybe it's going to make my life easier.
23:21
But to front load that like pain is kind of a hard ask.
23:26
You know that I like the mattress, but I think this is a
23:28
completely wrong way of thinking.
23:30
Yeah, All right. Well, I'm glad that you repeat
23:32
that. Why?
23:33
Because I think these tools are going to be used only if they
23:36
are immediately helpful to you. There's no way that you can push
23:40
these tools down the throat of people.
23:42
By the way, this is smart colleagues.
23:44
They know how to work and ready, they've been optimizing their
23:46
work, dealing with, let's say, inefficiency all the time.
23:49
It has to be beneficial to you, right?
23:51
And you have to be the one that recognizes that if that's not
23:54
the case, you should not do it in the first place, right?
23:56
So there's no sense to do something doesn't help right?
24:00
Now what we can do is that we need to make sure that the
24:02
barrier to get to that point is as low as possible.
24:06
That means that tools work. They give access to all the
24:09
knowledge bases that you have prevent you to make big
24:12
mistakes. So you got to say boundary
24:13
conditions provide examples, right?
24:16
That of what others are doing. So the the the barrier should be
24:20
very, very low. But once the barrier low, the
24:23
only thing that makes sense is that if you do, you create
24:26
something that helps you. If it doesn't help, you should
24:29
not do it, right? So yes, it's fine to test here
24:32
and there, but that's why we're not even thinking about
24:35
allocating special time for this should not be necessary.
24:39
It should be immediately useful, right?
24:41
So otherwise he actually create friction and then you have to
24:44
deal with friction. Then you have all the other
24:46
discussions. So we don't I don't, I don't
24:48
think it's a good idea. You also need to recognize then
24:52
the excellent work that some people do.
24:54
So one of the instruments, and there are many, is gamification.
25:00
So we have this program which we call Process as talent.
25:06
I think, I think, I think that's the name.
25:07
But anyway, it's it's designed around Shark Tank.
25:11
It's a competition. It starts now across the entire
25:15
group. Everybody creates agents, right?
25:18
And there's going to be, let's say, a committee that selects
25:20
agents every month. So the best ones are going to be
25:23
published across the group. They're going to be, let's say,
25:26
featured. Everybody knows about that.
25:28
And then there is this, let's say, crescendo to word smart.
25:32
There's a big event and there's a big selection of the winning
25:35
agent. And the winner will spend the
25:37
week in the Silicon Valley and many other things, right?
25:40
So there is a reason to do that because it makes a difference
25:45
for you, but it's also something that, you know, it is recognised
25:48
by others. So you need a lot of these
25:50
things, but the idea of separate, say, making this a
25:53
sort of separate program of innovation, I think that's not a
25:57
good idea. Yeah, I I like that.
26:00
And the real key there is you are now creating a very high bar
26:07
for the product itself, for the agent builder, because it has to
26:11
be instantly useful. It has to be immediately useful.
26:14
And we also accept, and we're really, I think it's a very good
26:19
idea is that let's say certain portion of these agents are just
26:24
experimentation. I'll do it because I really want
26:27
to stretch the tool and find out what it does, what it doesn't,
26:29
right? It's not a waste of time.
26:31
You know, you just call it learning.
26:33
And I think it is learning. Learning is not a waste of time.
26:37
Right. Yeah, of course.
26:38
And you might win a trip to Silicon Valley.
26:40
And then you might even get a trick to Silicon Valley.
26:44
That's so cool. But maybe it's a good moment to
26:48
shift a little bit towards the idea of governance and how
26:51
you're thinking about 30,000 agents.
26:54
You kind of mapped out, you have boundary conditions, but then
26:58
there's probably a lot of other pieces that you're thinking
27:01
about when it comes to the governance of how these agents
27:03
work and what you're thinking through as you're looking at
27:08
like, OK, we're introducing AI into the company.
27:11
What does that mean on a governance level?
27:16
That's an interesting topic. Governance in general governance
27:22
work when people don't feel that, but it is somehow
27:25
protecting them without putting so much pain or or or or hassle
27:31
to the people. So obviously it depends on, on,
27:35
on on the area. If for software engineers, the
27:39
governance is if it's generating code or reviewing code, you need
27:43
to have the technology or the tool sets that is basically
27:46
govern governing what's your agent is doing.
27:49
So there should be other agents or people governing what what is
27:53
going on so that it will allow the software engineers to be
27:57
more flexible, more confident about what they are doing.
27:59
So you need to build those things.
28:01
So you cannot just take governance as a Google form that
28:04
everyone needs to submit something before doing it or
28:07
ADPO needs to approve the use cases.
28:10
So that is bureaucracy that is also needed.
28:12
Don't get me wrong, for for cases that is needed as well.
28:16
But the idea is how we are going to build the governance so
28:19
people don't feel that there's a governance happening or it will
28:22
be quite practical for people to understand that, OK, they are
28:25
asking these questions or they have these boundaries or
28:27
limitations to protect me or the business from from these cases.
28:31
So in, in, in the early days of AI hype, we had almost no
28:41
governance in some in terms of tool set selection because
28:44
there's just so many tools available for marketing, for
28:47
sales and other stuff. And we have, I mean, it didn't
28:50
took more than six months to realise that everyone is buying
28:53
tools without talking to each other.
28:55
And even they're buying the same tool from different departments
28:58
with different contracts as well with different pricing.
29:01
But as as you all said, it is, it is fine because you also
29:04
don't want to slow people from from adopting or trying these
29:08
things. There's definitely a cost of it.
29:11
And usually the technology department pays the, the cost in
29:15
terms of making sure these things are secure or, or have a
29:17
proper procurement as well. But you just want them to try.
29:22
And after 6-6 months, almost 8 months, we started to put more
29:25
governance about, OK, how you can procure an AI tool, how you
29:29
can launch it, how you can make it access to these things.
29:31
But you also try to be practical in these ones.
29:33
You don't want people to feel like, OK, you are killing
29:36
innovation of the business. You are just slowing me down.
29:38
You're just making it harder for me to do just because you don't
29:42
have the budget. So we don't, we don't want to do
29:44
that, but we started with almost no governance.
29:47
We are right now in in the middle.
29:50
I don't know if you will go more on to the more strict governance
29:53
over time. I think our approach should be
29:55
more at automating these governing tools so that you will
29:58
feel less about someone is governing or something is
30:02
governing. So we build autonomy within a
30:05
framework kind of solution for, for, for the business.
30:09
And again, it depends on the area.
30:11
Software engineering is different, procurement is
30:13
different, launching a a sales AI agent is completely different
30:18
or a customer service to use cases completely different.
30:21
So as long as you are practical, I think you can find a solution.
30:25
And so if I can add a few things, because you mentioned
30:28
there, something I think is very important is it's really hard to
30:34
get these tools to work 100% right.
30:37
So you have to accept that they make mistakes, that they're say
30:41
their residual hallucinations still there.
30:44
And sometimes they, you know, they're, they're just not the
30:50
typical tool, which is deterministic.
30:52
If I have, let's say I follow a certain sequence, I'm going to
30:56
get always the same outcome. So these tools are not like
30:58
that, but that's, that's in nature of that.
31:00
So you need to accept a little bit that even for
31:03
experimentation, it can go somewhat wrong, right?
31:07
It's just the nature of that. So the question is, what is it?
31:10
The things that you accepted, what are the things that you
31:12
think they're not acceptable, right?
31:14
Well, you have many levers to decide that.
31:17
So the first one is, let's say whatever authentication
31:21
authorization these tools have, it should never be more than
31:24
what people have. It can be less, potentially even
31:28
less than you, but anyway, no more, right?
31:30
So that's one thing. The second thing that at this
31:32
point in time, I think we are sort of uncomfortable letting
31:37
agents do their work, completing autonomy.
31:40
So there's always a level of supervision and that supervision
31:44
is at the end of the process or it can be weaved in into the
31:48
process itself. For instance, we know that
31:51
prompt injection in many agents, it's still an issue.
31:55
So you want to make sure that if there is a sense that perhaps
31:58
another model says, you know, this is potential risk.
32:02
You have a pop up, you have something that the human has to
32:04
say. Do you really want to visit this
32:07
site? Or better, do you really want me
32:10
to go to that site right? Or execute this.
32:12
Or execute this. So you need to introduce a lot
32:15
of these checks in between and, which makes let's say the
32:19
process potentially slower. But in any case, you enable more
32:22
use cases in that, right? So a lot of these things are
32:27
part of the way you design things.
32:29
The other, the other observation is how do you figure this out?
32:34
Which, which are these measures, right?
32:36
By enabling testing, because people come up with additional
32:40
ideas or issues and threats and therefore you react to that.
32:44
So to the extent to which you try to limit the worst case,
32:47
right? So you really say what is the
32:48
worst case? What's the acceptable worst
32:50
case? Then you have the baseline on
32:53
top of that you can develop. But I think part of the
32:57
experimentation and part of the education to people using this
33:00
tool is that you have to accept 8020.
33:04
So there is they will at this point in time, you know, it's
33:08
experimenting. They not work.
33:10
They're not going to work to the the level you want.
33:13
You need to do some work until you get them to do exactly what
33:16
they want, right? But the experimentation has to
33:18
be based on this acceptance. Yes, it's going to be some
33:21
iteration before I get them to do what I would need.
33:24
Yeah. And you said something around
33:26
the sprawl of tools and just being able to grab whatever you
33:32
needed. And then you you kind of had to
33:35
tighten up the policies a little bit more on that.
33:37
It reminded me of a piece that my friend told me on how his he
33:43
was in a data governance role and his job was to go around and
33:47
check all of the different tools that were being used that were
33:53
with AI or whether it was like cloud code or it was a marketing
33:58
tool that sent data somewhere that was being processed by AI.
34:02
And he thought like, yeah, all right, cool.
34:04
We're not that big of a company. There's probably 10 or 12 maybe.
34:10
And after he went around to every team, it was like, oh,
34:13
there's 90. And that's not even all of them
34:16
that I know about, because I'm sure there's people that are
34:18
grabbing some data and then throwing it into their personal
34:21
Gemini account or whatever. And so in that regard, I, I
34:26
often wonder like, because it is such a tool that we want to be
34:34
able to like enable folks with, but at the same time you can
34:39
open yourself up to all these different vectors that you
34:42
weren't thinking about before. How do you think through that in
34:48
a way that is like, OK, now we are paying for 90 tools that we
34:53
didn't realize we were paying for.
34:54
And these tools, we may or may not be using them to send data
35:02
off like through another service, right?
35:05
So in a way, I'm thinking about the product management tool that
35:10
records your calls and then sends all of that, all your
35:13
calls to Chachi BT to get it analyzed.
35:17
You're paying for the product management tool.
35:19
But then you have to know in that governance position that,
35:23
oh, all of that data is going to ChatGPT.
35:26
Yeah, well. Practically, the problem didn't
35:29
start with AI, we had that problem in the past as well.
35:32
Still for for almost since technology started, every
35:37
department can buy a tool for themselves.
35:42
And they do. And then they do, yeah, like 2-3
35:45
years ago when the enterprise IT team showed me the number of
35:50
licenses we have from different tools paid by the credit cards
35:55
of the department deeds. It's mind blowing.
35:58
And they had no AI in any labeling or domain name.
36:01
So this problem always, always existed.
36:05
So what, what we built as a technology department is 1, how
36:08
we are going to basically secure ourselves because the, the
36:11
concern is that you, you understand, people want to make
36:16
their organization more productive and solve the problem
36:18
as soon as possible. That's, that's the
36:20
entrepreneurial mindset. So that's, that's still to
36:21
define, but we also have a responsibility to protect the,
36:25
the business and, and the customers and the partners as
36:27
well. So for that, we need to have the
36:30
policies that are practical so people can follow.
36:34
And at this even they buy something, they can just tell
36:37
that, OK, I'm buying this one. Hey, would you like to integrate
36:39
to our Okta system or our security system?
36:43
Is it behind Zed scaler or or something else And you need to
36:46
educate people about OK, we know you are building a shadow IT
36:51
from time to time, but at least let us help you so that we won't
36:55
have a bigger problem as a business.
36:57
So that's that's that's we have been trying to do and I think we
37:00
are more successful because the technology also helps about
37:03
understanding which tools are being used.
37:06
The finance reports tell us what is going on.
37:08
The network log says it OK, there's so much traffic going in
37:11
this in this direction. We try not to prevent it from
37:14
from the beginning, but we are trying to educate them to to to
37:18
help them help us protect them and the business as much as
37:23
possible. Because I think 111 concern I
37:26
usually have about the technology department or our or,
37:30
or me or my people is that we sometimes overestimate our
37:35
capabilities in terms of, or I can solve their problem, I just
37:38
need five more engineers. Or I can also solve marketing
37:40
problem. But that's, that's not the best
37:43
use of your five engineers for that problem.
37:45
There's only the solutions out there.
37:47
Let them just just outsource this problem to someone else.
37:50
We have a core business problem, food delivery, grocery delivery.
37:53
We need to use that five engineers to solve our problem.
37:57
We, if it's 100K that we will spend, let's spend it as a
38:01
business. So that's why we need, we need
38:03
to I think empower the people to basically operate in a, in a, in
38:08
a given framework and help us help them about how we will
38:15
govern or procure or or or maintain these things.
38:17
And I will be honest, I think we didn't nail it.
38:22
We still have painful areas as well.
38:24
But in order to 100% nail it, then I think you need to have a
38:28
very strict governance, which will eventually come with the
38:31
cost of innovation in the business.
38:33
So you need to take certain risks based on, based on your
38:37
business model, based on, based on what you have and educate
38:41
people to help you. That's that's how we try to
38:43
solve it. I totally agree.
38:46
And the, and so the education comes in many different forms
38:51
and so on. But one way you, you, you
38:55
already said that, right? As these tools, they make new
39:00
mistakes, but they can also make majority of mistakes we already
39:03
made before. They're just better at them.
39:06
The sense that you have more ways to do that, you have more
39:09
ways to make mistakes and so on. So you need to make sure that
39:12
people are aware of that, right? But you can also act at various
39:16
levels. For instance, you can act, let's
39:19
say at the point in which you say this type of data and
39:23
there's a very specific list of data types should not be used
39:27
with these tools, full stop. That's it, right?
39:30
So that's you cannot go there. Then you can also introduce a
39:35
certain number of friction in various places.
39:39
For instance, if you have an agent that sends e-mail, is not
39:42
allowed to send bulky mail, only one at a time, right?
39:46
And you are not allowed to loop it, right?
39:49
So that you reduce the chance that this is going to start
39:53
spamming the word. And so, so there are a variety
39:56
of measures that you know can be put in place.
40:01
And if you put them all in place, you eliminate all errors.
40:04
And then you don't really have much use of these tools, right?
40:08
But then you start peeling them away to see what's acceptable.
40:11
It's not acceptable. You get to the point in which
40:12
you you find, let's say, some sort of balance between the need
40:17
of innovation and also the need to protect the organization.
40:20
Yeah, that's for sure the case. What is it's, it's also
40:23
extremely important is that lot of the things we learned in the
40:26
past about all this are still valid, but these tools are also
40:29
new. So you find out a lot of the
40:33
vulnerabilities if you want by doing right.
40:37
Therefore, if you have tools that enable you to test without
40:41
making big mistakes, then you find the things that can go
40:44
wrong, and then you serve that and then on and on and on and on
40:47
and on. It's a process of learning the
40:49
fastest. You learn the fastest.
40:51
You can use these tools at scale for important things, which is
40:55
very beneficial. I want to end with one question
40:59
that wasn't on the docket, but as we're talking, it's like it's
41:04
very important and I've heard folks in the community talk
41:08
about it. I've heard folks in different
41:09
communities talk about how they are thinking through
41:14
provisioning resources and provisioning mainly finance
41:19
towards AI because it is such a like we saw the graph yesterday,
41:25
it was exponential. Did you know it was going to go
41:28
exponential? Probably not.
41:29
Were you thinking maybe it could?
41:31
Yeah. But how are you allocating
41:34
budget towards GP, US or Open AI bills or all of the different
41:40
things that come with the AI if you are going to see the
41:43
explosion that you're expecting? So that's a super interesting
41:47
and important point too, because we we see two phenomena.
41:50
On the one hand, let's say the let's call the unit cost of
41:54
intelligence goes down very quick, right.
41:56
So we're doing the same thing a year ago and this half and half
41:59
and half and half, right, the token, the cost per million
42:02
token is also going down. So there is a trend in our
42:05
favour. And all this cost will be let's
42:07
say decreasing over time. At the same time, we see another
42:11
phenomenon is that everything that we do requires more tokens,
42:15
it costs more. So these two things are one
42:17
against the other all together. It goes towards being more
42:22
expensive, right? So the additional use of tokens
42:25
is overcompensates the reduction of costs.
42:29
At that point you have many things that you can do.
42:33
One thing is optimization. Many use cases developed so far
42:38
are super, let's say inefficient.
42:41
The prompts are very large, the contests are very, let's say
42:45
wordy and all this kind of thing.
42:47
So you you need to go back and redesign how things are done
42:51
because otherwise, you know, it costs and so on.
42:54
There are many tricks and many heuristics to reduce the number
42:57
of tokens at the output for many use cases.
43:00
So people start getting, let's say, paying a lot of attention
43:04
to optimization of this. You start getting to the point
43:07
in which open source models are competitive to with, let's say,
43:12
the bigger commercial models, not for, yeah, so not for
43:16
everything, but for many practical use cases.
43:18
Then at that point, you're just better off hosting the whole
43:21
infrastructure yourself, right? Which is another way of covering
43:24
it. But All in all, it is it was not
43:28
a problem last year because volumes were small.
43:32
And then let's say use cases were not complex.
43:34
It is an issue now. That is exactly what I was
43:39
thinking because especially as you're scaling much more and
43:43
you're expecting that usage is not going to slow down anytime
43:47
soon. No, like that scale is only
43:48
going to increase. You're going from 18 to 30.
43:51
I imagine the next goal and that's just until March, by the
43:54
end of next year, it's. 100,000 much more complex and then let's
43:59
say the other thing which is obvious now is that the
44:01
complexity of use cases comes with the with the increase in
44:05
costs, right? In the past, independent of the
44:08
the complexity of the question or the task, the usage was the
44:12
same. Now there's a correlation
44:13
between complexity and usage, which also good from the
44:17
business point of view because you solve a bigger problem.
44:20
So it makes sense. But still you need to, to, to,
44:23
to, to start paying a lot of attention because it becomes a
44:27
significant operating cost if you let it go.
44:30
Yeah. Yeah, I mean, I, I agree.
44:34
I think, I think deep inside you believe that's the return of
44:40
investment of this is always positive somehow.
44:42
Like maybe not in the beginning because the learning curve of
44:45
wrong, wrong agents are being billed or or tools are put.
44:48
But over time with people learning, with the technology is
44:52
getting better, the adoption getting higher.
44:53
You know that the return of investments is going to be
44:56
higher for the business and every business wants to grow
44:59
more. And I think we all believe that
45:03
AI is going to be the game changer of how fast the business
45:07
can grow or how fast the business can fail.
45:10
So do you really want to try to optimize your costs on these
45:15
ones or make sure that this company will get 2 times bigger
45:19
in five years time or will shrink half in five years time?
45:23
That's so that's, that's the, that's the concern.
45:25
Obviously you will optimize those things over time, but I
45:30
think truly we all believe that the return of investment will
45:32
always be higher for the business if, if you will make
45:35
people use these technologies much better than they have been
45:39
using a year ago or two years ago.
45:40
So, so you need to have conversations with your CFO or
45:45
or other leaders to, to, to manage that.
45:49
But at the end like if people are using it, they are finding
45:52
value out of it. No one is just inherited trying
45:55
to waste money of the company. They get more out of these
45:59
things and then the company spends.
46:01
Obviously probably some of them don't have the most optimized
46:04
prompt or or the right model. But I think the these are very
46:09
small unit cost at the end. That is not the biggest concern
46:12
of me at the moment. My, my concern is how we are
46:15
going to grow the company to three times in five years time
46:18
and how AI is going to enable that.
46:28
Something funny happened to me six months ago.
46:30
I asked Deep Research for a report on different GPU
46:33
providers and it was absolutely shit.
46:37
I couldn't figure out what each Neo Cloud Value prop was, so
46:42
that set me off on my latest side quest of creating a
46:46
practitioner's guide to choosing GPUs.
46:49
I'm happy to announce this guide is now ready to see the light of
46:53
day and you can download it for free right now by clicking the
46:57
link in the show notes. We've already got some community
47:00
members feedback of what they wish they would have known
47:04
before signing a gigantic contract, and I would love to
47:09
hear from you if this provides any value or you have some
47:12
things that the rest of the community should be thinking
47:16
about when they're on the market for GPU's.
47:18
Go ahead, download that resource right now, completely free.
47:22
It's in the link in the show notes.