Latest / Agentic Conversations (formally mlops.community) / What is AI Quality? // Mohamed Elgendy // #228

Transcript

What is AI Quality? // Mohamed Elgendy // #228 · Agentic Conversations (formally mlops.community)

0:00
Hey, I'm Mohammed, CEO and founder of Colina, and I'd like
0:04
to drink my coffee black with two Splenda.
0:07
What is up ML OPS community? We are back for another podcast.
0:11
I'm your host as years rolled Dmitri Ose and today we're
0:15
talking to Mohammed, my man, my close friend who I have been
0:18
collaborating with a lot recently.
0:21
We are doing this AI quality conference in San Francisco and
0:27
he is the other Co organiser, the other brain behind it.
0:30
And I've got to say, I had to have him on here just because
0:34
his vision for AI quality and what it means that the industry
0:39
is lacking right now when it comes to putting AI in
0:43
production is so cool to see. He breaks it down.
0:47
I went deep on how you can look at testing, how you can look at
0:54
quality, how you can look at frameworks.
0:56
And one thing that I really appreciated from this
0:58
conversation was how he broke down the gold standard that
1:03
we're trying to create here with the conference.
1:05
And after the conference has finished, we will continue to
1:10
have working groups creating AI quality gold standards.
1:15
And he breaks down what that means, what the working groups
1:18
are going to look like. And we're launching it right
1:21
now. If anybody is interested in
1:23
joining one of these working groups hit us up because
1:26
although there's still about two months until the conference
1:30
happens, we're gonna get cracked and the wheels are in motion.
1:36
So last but not least, I mean he he dropped so many gems on me
1:41
that I I really want to say everything, but I don't want to
1:43
spoil anything. So let me see if I can be vague
1:45
enough to not kill it for you when you hear him.
1:48
But his idea of making sure there are standards, that for
1:54
each industry you have standards.
1:55
And then inside of each industry you have for each application
2:01
you have these gold standards. And what these gold standards
2:04
look like is what I really enjoyed because you can easily
2:08
throw around the word standards. And I actually grilled him on
2:10
this when it comes to how are we going to make sure if we create
2:14
standards, that doesn't just become another meme like the Oh
2:18
yeah, we had 12 standards yesterday and now we have 13
2:22
standards today. That comic that I'm sure
2:24
everyone who has thought about creating standards has faced,
2:29
and I thought his answer to that was insightful and brilliant.
2:33
And so this idea around industry and applications inside of the
2:38
industry and each of those having their own quality
2:40
standards was so cool to see. I really enjoy the way that he
2:45
thinks of things and how he's moving the ball forward in this
2:49
domain. Hope to see you in San Francisco
2:53
June 25th. If you can't make it hit us up,
2:55
we would love to chat with you all about these AI quality,
2:59
metric, standards, testing, You name it, we got it.
3:03
And as always, if you like this, share it with a friend because
3:07
you know the world needs more AI quality and especially those big
3:11
businesses that are losing money every time they put out one of
3:14
those shitty chat bots. Alright, see ya dude.
3:23
Ohh, we gotta start with the question that I know everyone
3:27
wants to hear in your eyes. What does AI quality mean?
3:35
What is the AI quality? So that's is the, that's the
3:38
question of the hour, right. Everybody's asking this
3:40
question. We talked to hundreds of
3:42
customers all the time. And then the question is and
3:44
even when we talk about more details about let's build
3:47
quality standards, people are asking what does that mean.
3:51
Let me actually tell you a a a good story here and then we jump
3:54
into it. So if you give context for
3:58
everybody as well, we at Colina, we're building AI testing and
4:01
quality validation tools. So we've been working with
4:05
enterprises, enterprise customers and leading startups
4:09
building on top of general I believe in the products.
4:11
If we're talking about the kind of what has what became all the
4:14
right now and then we always think about OK, what is this,
4:18
what can we do more right And we are testing solution and we see
4:22
our customers testing and rigorously testing their models
4:24
and finding exactly where the failure points look like.
4:28
But we still feel we have felt that something is missing
4:31
something in in their testing process or their validation or
4:35
their seek for high quality product, something is missing
4:39
there. So we were having one of our
4:42
internal product design sessions and then our CTO and founder and
4:48
Rishi shared like a really good insight like he asked the
4:51
question through through a statement on the table and they
4:53
just kind of changed our perspective.
4:56
He mentioned, he said we are providing testing to our
5:00
customers, we're not providing quality.
5:02
And that had us start thinking about the question that you just
5:05
asked, what is quality, right. So we started thinking about,
5:08
OK, how do we, what is that, what is quality and how we can
5:11
define that and actually help our customers achieve because
5:14
that's the end goal. Testing is a step in the quality
5:17
process towards that quality goal, right.
5:20
So we started zooming out and thinking about what is AI
5:23
quality in our case. And then the quality here is
5:27
dependent on the application and the domain that this application
5:31
is being used or being the problem that's being solved, to
5:35
give you an example here is just outside the AI site and we come
5:38
back to it. If we're thinking about like
5:41
this band here, it's $1.00 pen. And is this $1.00 pen better
5:45
quality or would this pass the quality bar more than $100 pen?
5:50
Now this is a quality question, not has nothing to do with with
5:55
AI but it's a manufacturing quality question and they had
5:58
this question asked the same question what is quality 100
6:01
years ago when they are building these standards.
6:03
So the answer here is this $1.00 pin can pass the quality bar and
6:09
$100 pen not pass its own quality bar.
6:12
And this is to answer questions. There are different quality
6:15
standards based on the application that you are using
6:18
it for. So the $1.00 pen here is not
6:20
expected to perform under severe weather conditions, for example,
6:25
right? It's not expected to live for
6:27
years, it's expected just to be disposable, Use it for a couple
6:30
of months and then throw it away.
6:31
Now having that context defined, let's think about Jenny I for
6:35
example, if you're building a chat bot for that will be used
6:39
in Twitch for gaming industry. Now the quality centres for that
6:44
is completely different than the quality standards from a chat
6:47
bot used in a fintech company or in a bank where you are
6:51
prohibited by law to give a financial advice.
6:54
And even if you're allowed to give a financial advice, you
6:56
have to be very precise. You can't say three years, the
7:02
price value is going to be 3 years.
7:04
It's always gonna go high. You have to be very precise in
7:07
timing. And actually what is that
7:10
change? So with that then gold standards
7:14
or quality standards for the AI product.
7:16
Now in our definition, it means it's built for, it's built for
7:21
specs for the specifications that is meant for and deployed
7:25
safely. So that's the whole thing like
7:27
hallucinations and all the issues and risks that come with
7:30
it, this is these are all called risks like NIST, the National
7:34
Institute of Standards and Technology in the US they define
7:38
that as risk management framework.
7:40
So if I am a data scientist, I'm thinking about what is EI
7:43
quality for my product. It is that is 2 main sections. 1
7:48
is what are the specifications, what is it intended to do and
7:53
what are the risks that are associated in your industry.
7:56
Then you're able to benchmark for test your models or your
8:00
products based on that. So we ask a lot of the, then I
8:04
will give you the short answer, right, like what is the high
8:07
quality? We asked a lot of leaders in the
8:10
data science and AI space and then they tell us the name and
8:14
terms, which actually makes a lot of sense.
8:17
The, the quality means that my product performs as intended,
8:21
right, safely, that no risks associated with it that we don't
8:25
know about. And that is the fact that
8:26
happened to us to start thinking, OK, for us to deliver
8:29
quality to our customers and to the industry, we need to define
8:33
these gold standards for each domain.
8:36
Brilliant. There's so much to unpack there,
8:38
man. Especially, I really like this
8:40
idea of the pen and you have two separate pens.
8:44
One maybe has diamonds on it, the other doesn't, and you have
8:48
two separate quality standards that when each one of these
8:52
comes off of the assembly line, they're going to be inspected
8:56
and they're going to pass the quality check.
8:59
If they or if they can pass the quality check, that doesn't mean
9:04
that the $1.00 pen is going to be able to pass the quality
9:08
check of the diamond laced pen, right?
9:11
Absolutely, yes. So and then if it's Diamond
9:13
Lace, you'll have your Your quality requirements will
9:16
include the diamond quality requirements too, because it's
9:19
the quality requirements for your product.
9:21
Yeah, exactly. So there are many different
9:24
layers of that quality to make sure that it's not only the
9:27
actual ML, but then you have the software, then there's that
9:32
tried and true practise of yeah, quality assurance, and then
9:34
you've got the actual product. Yeah.
9:36
If it's sitting, let's say, if you're deploying on Edge, right,
9:38
then latency is part of your quarter report.
9:40
I mean, if it's not like it, but in terms of latency is in fact
9:44
much more if you're deploying on Edge on some hardware.
9:46
So you're thinking about the specs of the hardware should be
9:50
that high or that good. So it gives you a response
9:53
within half a second or 500 milliseconds versus a couple
9:56
seconds if you're doing on cloud for example.
9:58
So that could be part of your quality requirements when in in,
10:02
in the on edge deployment or on time sensitive applications.
10:06
Depends on how we start thinking about breaking down these
10:09
standards. Well, you mentioned gold
10:12
standards and knowing that each one of these specific use cases
10:18
feels like it has a new set of gold standards, how do you go
10:22
about creating that? Because if I'm thinking about
10:25
it, I'm like, OK, cool, the financial Advisor chat bot.
10:30
It's never going to say that it's giving financial advice,
10:33
but that's one lane of gold standards.
10:38
I feel like you have to really curate and create gold standards
10:43
for that use case. Is it like you need gold
10:47
standards for each specific use case or can you generalise the
10:50
gold standards like breakdown what exactly these gold
10:53
standards look like? Yeah, so it it will not be too
10:57
too detailed to the use case level.
11:00
You can generalise that more that provides the same details
11:04
but it doesn't have to include the overrunning work of defining
11:07
each specific use case. So it would be for the domain
11:10
like let's say contact or banking and then then within the
11:15
domain you will have to fork out for what kind of application
11:19
it's been used in. It could be in in banking but
11:24
you still it's a customer support kind of banking
11:26
application or in banking as well in financial services.
11:30
So if you are seeking let's say Fast forward many months from
11:34
now, if you months from now and we have like a rigorous gold
11:37
standard developed by the industry.
11:39
You probably you need to get into OK Jenny I standards and
11:42
then double click into that and then get inside the financial
11:46
sector or the banking industry and then get into that And then
11:49
you're building a customer support but you will see it
11:52
there if you get to see what are the functional firms that should
11:57
be thinking about and the acceptable limit.
12:00
So we talk about latency just talked about it now then what is
12:04
the acceptable limit here for a bot that is a customer support
12:08
It's maybe it's fine to give you an answer in five or six or
12:11
seven seconds, right And then you can add a spinner or
12:13
something that for UX. So and then that's one thing and
12:17
then you go back to the financial sector again up and
12:21
then you go down into financial advice or financial advisor.
12:25
Now you have different set of requirements that you you need
12:28
to think about and they're always would be thinking about
12:32
this for months. And then there are almost always
12:34
fall under one of two categories. 1.
12:37
The functional requirements which is what you're attending
12:40
your users to use which is the use cases that you talked about
12:42
and the other ones are the risks associated with the technology
12:47
that you're using. The technology that you're using
12:49
here in this case is Genii. Now what are the risks?
12:52
We all say hello, what are these hallucinations right now?
12:55
If you alienation that could be here.
12:57
Like we said, like how, how? If you're asking a question and
13:00
it's not part of your right system, you don't have the
13:02
answers for it. It can confidently give you a
13:05
wrong answer. That is the known type of
13:07
hallucination. But another type of
13:09
hallucination is jailbreaking. So it gets outside the
13:12
guardrails. That was built around that.
13:14
It was built around like actual builders of the application like
13:19
don't give financial advice or you can give financial advice
13:23
pre empting it with some kind of a caveat based on my
13:26
understanding of the industry and and and else is here is what
13:31
I can tell you as an automated bot, but there are some
13:34
regulations here. This is when we were at work
13:36
with regulatory bodies as well to tell us how we can bridge
13:38
this gap. These standards have to be
13:41
defined for the industry and for the type of application.
13:44
And then now you get to say, OK, the acceptable score because now
13:50
we work with everything you said.
13:51
You need to figure out the metric for it and then how to
13:54
test it and what's the acceptable score, because you
13:56
get some scores at the end and you don't know what is that
13:58
score for, right? You don't know if that's good or
14:00
not. So based on the acceptable here
14:02
are the acceptable limits from 70 to 90% kind of accuracy in
14:07
these things based again on the application that is being
14:10
developed? So it feels to me like and and I
14:14
I appreciate the idea of industry and application because
14:18
it feels to me like you have a certain baseline that you're
14:23
probably going to see across the majority of the applications and
14:28
industries. So if you have a chat bot, a
14:30
financial use case, it's probably going to need a lot of
14:34
the same stuff that a chat bot in the insurance use case needs.
14:38
Or if it's a customer support in finance, it's probably very
14:42
useful to have customer support for like a telephone company,
14:46
right? Absolutely.
14:47
Then you depending on each one of these specific verticals and
14:53
each one of these industries, you're bolting on specificity I
14:57
would imagine. So what is wild to me is how you
15:01
need to fan out and think about all the different use cases and
15:06
then all the specific risks that are inherent with these use
15:10
cases and in these industries. Yes.
15:13
So it is a huge undertaking here and it doesn't, it doesn't take
15:18
one person or one company or one institute to go ahead and build
15:21
this whole thing. And that's why you know that we
15:24
are, we are working together Colina and community and hosting
15:29
the air quality conference and the AI Quality conference which
15:33
we're thinking about it here as we try to find another word
15:37
other than conference. But this is the kickoff of this
15:39
air quality movement. We are putting together a nine
15:42
month programme to build these quantities, standards and to be
15:47
able from our understanding and we believe that the best way to
15:51
build a applicable and safe gold standard for quality, you need
15:57
three main layers to be in the room, the builders, the AI
16:02
builders, we understand that and regulatory bodies obviously
16:06
there and the apps or the tooling and the infrastructure
16:11
layer. So I've seen like the quality
16:13
triangle that these are the three diagonals effect of it.
16:18
There you go, Love it, because these are the most OPS toolings
16:21
or the infrastructure companies. These are the people who are
16:24
going to make it feasible, automatable because the builders
16:27
cannot be installed down or slowed down by a lot of
16:32
governance work or providing a trace to their data and their
16:37
models, everything that's required.
16:38
But for that to be implementable and actually the AI builders
16:43
adopt this technology, it needs to be all these regulations.
16:46
It needs to be automated in a in a very supporting fashion to the
16:53
data scientists like they don't need to think about what's
16:55
happening in the back end. So we're bringing these three
16:58
players together in what we call the AI Quality conference, AIQ
17:02
Con. It's happening in San Francisco
17:05
on June 25th and in there we are, this is the kickoff and
17:09
then we bring in industry leaders, regulatory bodies and
17:13
obviously the infrastructure companies and teams and we are
17:18
putting together a nine month programme.
17:20
This nine month programme has three main phases.
17:23
The first phase is the, the, the initial phase like you're
17:26
getting to understand OK discovery, what are the domains
17:29
that we're thinking about. And we initially started about
17:32
three main domains, the robotics, robotics we talked
17:37
about now it's called ALDI, but the robotics domain think about
17:43
all the robots including Thomas vehicles that are out there
17:47
being deployed, fun of us that we'll start seeing them.
17:50
So that's that number #1 do mean that we're focusing on and for
17:54
Jenny I specifically we're focusing on ecommerce and the
17:57
financial sectors, both of them are.
18:00
This is where we see the highest adoption for Genai and we're
18:05
working on defining gold standards for those.
18:07
So that's Phase #1, the discovery phase and we will have
18:10
three months for that. And then we have the second
18:13
phase is we want to define the risks and the application
18:16
guidelines, specification guidelines.
18:18
So for each of these two means to be able to allow the provide
18:22
some kind of portal where people would go ahead and selected
18:24
domains like we mentioned and then zoom in to really quickly
18:28
what is required from them, what are the risks expected.
18:32
And this has to be something that is always updated with the
18:34
updates and is still in decent stage now.
18:38
So this is growing with every all the releases that we see
18:40
every day. So second stage is defining
18:42
these risks and the specification guidelines and the
18:46
applicable or the acceptable scores or guidelines for these.
18:51
Now last thing is the third stage which is the last three
18:54
months. We want to define the processes,
18:56
the tools, the infrastructure makes something that is
18:59
applicable. We want to see industry users
19:02
being able to just take this and give them to their teens and say
19:06
OK, start building against that. That would be a great stepping
19:09
stone for regulatory bodies to start thinking of a practical
19:13
way to enforce these regulations without stifling innovation for
19:18
the builders. Yeah, this is cool because it is
19:21
something tangible that comes out of a fun party that we're
19:27
having. Not exactly exactly.
19:30
We're going to have a great time and nine months later there's
19:34
going to be a baby that's born. It's not the baby that you think
19:38
is being born. The analogy here, I didn't think
19:42
about it. It's actually, yeah, it's
19:43
exactly 9 months and new baby was born.
19:46
That's a great analogy here. There we go.
19:48
And I really like that because the ability for us to be able to
19:53
say, all right, we're going to understand deeply from the
19:58
subject matter experts, from the builders, from the people that
20:01
are creating the tooling for this to all try and work
20:04
together and make this whole trajectory a whole hell of a lot
20:07
easier. Then that tangible thing that
20:11
comes out in nine months is going to be so valuable for the
20:16
community. Yes.
20:18
And so that's the goal, right. We we have been we we see
20:22
regulatory bodies are trying to protect the people us right
20:27
where we are the builders, but we are the people to be
20:29
protected. So they are rightfully doing
20:31
what they are trying to do there.
20:34
But again they are the AI leaders or industry the builders
20:38
they have the right to think about, OK, there's this tension.
20:41
I see it as a healthy tension, right, that's that's needed in
20:45
these kind of conversations and that's why I see that the
20:48
infrastructure teams are within organisations or companies.
20:52
They are they, I believe this is the the magic, you know, magic
20:56
stick here, like not necessarily like they're going to magically
20:59
make it happen. But more importantly, OK, how
21:02
can we use engineering to make sure that this is done safely
21:07
without stifling innovation And this way we are able to unlock
21:11
this bottleneck where everybody's standing in, OK, we
21:14
want to regulate you and OK, don't regulate us.
21:17
So in the in the air quality conference we are bringing, we
21:23
focus on bringing for when we're thinking about the leaders, the
21:26
builders, we focus on bringing the actual people who are going
21:29
influencing their teams. You can always get speakers
21:32
which they are great conferences to bring speakers.
21:34
But in in this specific conference we brought the heads
21:38
of AIS or the CTOS of these companies that are they are
21:42
coming part of a mission coming to kick off this mission with
21:46
all of us and everybody has 30 minutes, 45 minutes that's not
21:49
enough to you know share how they think about quality.
21:52
But it's great to start here is as a kickoff here is me as a
21:58
company how we're thinking about what standards here's so on
22:01
regulatory bodies are coming too.
22:02
So on the industry side we'll bring in Motion network for
22:06
example right Motion now we hear is the CTO and President of
22:10
Cruise. He's the man in charge of
22:13
bringing this technology out to the to the streets for us.
22:16
So he will be sharing cruises end to end AI philosophy and how
22:21
they are thinking about building new quality standards for
22:24
autonomous vehicles. Yeah, on the other desk for the
22:27
towns on the other end. On the other side we have
22:29
Richard Socher. He's a Stanford researcher, CEO
22:31
ofyou.com and he's sharing with us how they are building quality
22:37
standards to build trust and LM based applications.
22:41
Now Speaking of the of regulatory bodies and the
22:45
government, we have a government panel that's moderated by the
22:48
Washington Post. They are hosting Nest, they are
22:51
hosting the DVD, the representatives from the White
22:55
House who wrote the the executive order.
22:59
We also are bringing on the VC panel that is moderated by the
23:04
information reporter. So they are bringing in PCs or
23:09
investing in tonnes of AI companies and applications to
23:13
think about OK, aware how we can put our and deploy our resources
23:18
and our funds to promote this AI quality movement.
23:22
So many industries are coming like we talked about could you
23:25
probiotics store robotics that's an autonomous track on the land
23:29
track or generally there's opening eye, coherent topic and
23:32
many more. So there's tonnes of
23:33
conversations that we the goal of these conversations, every
23:36
leader is coming to say here's how we think about the equality
23:39
and here's we could we're going to work together to define these
23:41
gold standards. I just gotta say I am so excited
23:47
to wear my eye hallucinate more than chat GT.
23:50
Shirt. With a bunch of these government
23:53
folks that are making regulations, we're going to see
23:56
how that goes over and maybe some of them we're going to,
23:59
we'll bring some and I'll give some.
24:01
To that, yeah. I'm excited for this one.
24:03
I'll be. Ready.
24:04
By the end of it, everybody's going to be wearing that shirt
24:07
and it'll be a good time. It's a big step up for us, man.
24:12
Like, let's be honest, having the panel moderated by the
24:16
Washington Post for me, I never would have thought four years
24:20
ago that would be happening with the ML OPS community stuff.
24:25
So it's brilliant to see that. I also kind of wanted to dig
24:29
into this idea of we talked four years ago and you were one of
24:35
the first guys on the podcast back in the day and you were
24:40
talking about this. It feels like you've learned so
24:44
much in the past four years because you've been out there,
24:47
you've been talking with people, you've been recognising like
24:50
what people are saying as far as the quality standards, how they
24:53
need to go about it. And the first thing that comes
24:57
to my mind is that with this nine month project and we're
25:05
really going to try and create these gold standards, right?
25:09
How do we not become the meme of Ohh?
25:13
There's 13 standards, but none of them work.
25:15
We should make a standard that encompasses all of these and
25:19
then next thing you know, there's now 14 standards.
25:23
Yeah, I mean that that's a very good question.
25:26
We started thinking about this given that everybody is thinking
25:29
about gold Sanders differently and it's a challenge that we
25:32
will go through and to say up front it's an evolving process.
25:36
So there will be let's define standards first.
25:39
So if standards are list of like the way we envision what the
25:43
senders would look like, it will be the list of functional specs
25:49
was intended to do and potential risks for this domain now comes
25:56
with it the the methodology of testing.
25:59
So for each one it has a different allows you to test if
26:02
you're testing data leak. This is different than testing
26:06
for correctness or factfulness or the results, right.
26:10
So it has different methodology to test.
26:12
So we talked about the functional requirements and
26:16
risks, what's the methodology to test?
26:18
What are the metrics that you're going to use and what are the
26:20
acceptable guidelines. Now having these four, that is
26:24
the layout. This is the gold standard of the
26:26
gold standard, if you will, right now as people start
26:29
thinking about, OK, I'm gonna create different goal standards
26:32
using this layout, this framework of thinking, right.
26:35
So that is the framework. So that's one point we get
26:38
there. We're already winning as a
26:40
community because we're talking the same language, so that now
26:42
we're just agreeing or disagreeing on probably the
26:45
acceptable ranges. This is where things are going
26:47
to be left to that provider or maybe what are the metrics.
26:50
This is evolving but as long as we agree on OK, you have to
26:53
define the risks of the methodology to test that.
26:57
And so, so we have seen when we started, like you said we were
27:01
3-4 years ago when we started talking, we had the conversation
27:05
and I we we believe like we when we started there was Jenny, I
27:09
wasn't out there and we were looking obviously we believe in
27:12
their testing and quality for AI in general.
27:14
But when we looked at who is doing this too, we were the
27:18
first ones to start thinking about it and it felt like OK,
27:21
this could be something wrong, right?
27:23
Like OK maybe we're working on something that is not important
27:26
or is not feasible. You know you understand this
27:28
when you're working on a new startup or think about something
27:31
but at that was that that concern in our hearts.
27:36
When we started launching the company just went away in the
27:39
first couple of weeks talking to customers and before January the
27:44
amount of traction and that we got from customers like yes we
27:48
want testing and the biggest thing we heard that's an old AI
27:52
Ohh Mohammed, I thought we were doing it right now I say I'm
27:55
rethinking and that all that has been a challenge in early stages
27:59
when we are like in our sales or go to market motion that when
28:02
you message Demetrius say man you're a great scientist or
28:06
engineer. We are rigorous testing
28:08
platform. Always the answer is we have
28:11
something right? Because yes everybody has
28:13
something nobody is pushing out of Park And we realised that OK
28:17
the industry needs testing, needs rigorous testing to be
28:21
automated to right and but needs education on what is that
28:26
testing look like and why testing will save your effort.
28:30
If you're thinking about effort on upstream processes like data
28:33
labelling and training, you can just label only data that you
28:37
want. Saves 90% of your effort on
28:39
labelling instead of turning the the model, staying in the
28:43
experimentation stage and trial errors where CAMI informing the
28:46
model in these areas or not. Now we will exactly know where
28:49
your models are failing exact scenarios and then you go ahead
28:52
and makes them. So it turns the experimental
28:56
nature of machine learning into an engineering discipline where
28:59
you find a bug, you and test it and improve it and test.
29:03
So that was for what's neural network is and predictive AI now
29:08
Kim and Jenny, I probably a year and a half ago or a year ago now
29:12
they have been now the evaluation this called LM
29:16
evaluations now which is we like to think about as testing and
29:18
quality more has become the everybody's thinking about it,
29:22
right, which we're happy with this like it's not about
29:24
thinking about the competition. The business is big and the
29:27
industry is evolving very fast but it's more important like OK,
29:31
there's awareness here that we're in this part is gone but
29:34
we we found OK, it's still some awareness is needed.
29:37
Look at least awareness in terms of the the testing is needed,
29:40
quality is needed. Now the second level of
29:43
awareness is OK, what's the methodology what are the what's
29:47
the quality framework. NIST likes to call it risk
29:50
management framework. Europe called it the EU's
29:53
trustworthiness I guess act so or something that we use the
29:57
term around trustworthiness reliable or safe I something
30:00
like this. So it is all all falls under
30:03
quality and things that we been developing very fast since you
30:06
know I talked spent three years, but in the last 12 months every
30:10
month or so like you said, we see a new development and new
30:14
evolution in the technology and the standards in the
30:17
regulations. And we're working on this, we're
30:20
working on with everybody, trying to just make sure that
30:23
everybody is working together in defining these for the better
30:26
for the entire industry. You said something before that I
30:29
also wanted to dig into because you mentioned how testing is
30:34
just one step along the way to quality.
30:38
Yeah, I mean there, there are two main sections for this.
30:41
One is test coverage and one is what's acceptable, right.
30:44
So let's break it down a little bit.
30:46
One is test coverage, right? When you are testing, when
30:49
you're like software, right? When you're testing your
30:52
software, you have to assert a lot of functions, which is this
30:54
is what the unit testing would look like in software.
30:57
But you know the logic, right? You know what, you got it for
31:00
Testing and machine learning or AI is different.
31:03
You're testing based on datasets.
31:04
So when you are testing your model, here's a good example
31:10
actually outside LMS there, let's say generative AI, images.
31:14
Gemini, right. So Gemini is a great example for
31:17
you. And then exactly like you
31:18
mentioned, things are good until they are not going to.
31:21
Push something? Yeah, Worlds.
31:26
So that is exactly, that's a great example of, OK, how do we
31:30
push out our first product with no embarrassing risks or even
31:34
worse, no impactful risks, right.
31:38
So that's the first piece. And then obviously after that,
31:41
how do you, that's the first product, How do you keep pushing
31:43
out new releases and giving your customers the better product?
31:47
You want to see regressions and you'll keep capturing detail.
31:51
Yeah, because one of the things. Too right?
31:53
Like you update and then all of a sudden it's way first then the
31:57
last version. I know that has happened plenty
31:59
of times with GT3 and GT4. It's like, wait a minute, are
32:03
things worse? Now, yeah, it was good.
32:06
So that that's that's exactly the the nature of testing today.
32:10
An AI jar, well, there's old or new AI if we were to use these
32:13
terms. So if you're thinking about
32:16
Gemini now as a as a case, right, and then you're you're
32:19
writing usually testing is done. The old school testing is you're
32:23
you have some benchmark datasets of prompts or or or even data
32:29
and then you're just let's say the benchmark is like a million
32:33
data points and then you're running your inferences on your
32:36
train model and then you come out with OK my model is 90%
32:40
metric right F1 score, recall, accuracy, hallucination.
32:44
However it is right. So now this is 90% doesn't tell
32:48
you what's in the coverage in the test.
32:51
You have a bunch of data, usually you have some specs you
32:54
keep adding to it, but it's just a bucket of data that doesn't
32:58
say exactly what shows like. And then the second point is
33:01
the, the 90% metric, what does that mean right, Is it better
33:03
than 85%? We don't know.
33:05
Right So and so when you're setting up your benchmarks, it
33:09
has to be or the way we think about it.
33:11
Here is the test case based evaluation testing scenarios.
33:16
So you create your data, you certify your data or into our
33:21
products into specific scenarios.
33:22
These are the reflect your functional requirements.
33:25
So if we're talking about the case had a great and noble
33:31
goals, if you will, right, like the goal was was right.
33:33
Like hey, we want to diversify. We don't want this to have any
33:36
bias in race, bias in gender, that is great.
33:39
So that's the requirement. Now you need to the way you've
33:42
instruct the test case, you have to add the test examples and we
33:45
add that in our tools. That's why I'm saying you know,
33:48
infrastructure at teams are very crucial in this process because
33:52
they can just build in these cheques, data, cheques inside
33:57
the platform. So if that goes on to Colina
33:59
then the test case is for to verify diversity and no virus
34:06
right. So the data cheques should look
34:08
into negative examples too. So when you say who are the
34:13
founders of the United States you that the the quality check
34:18
here should show that OK you need for factfulness you need to
34:21
deliver the right it's necessity right like that's that's the
34:25
critical part because it's historical event.
34:27
So in that here that hasn't been the test coverage wasn't tested
34:31
that and then I don't know obviously what's happening
34:34
inside how they are evaluating Jim and I.
34:35
But after after that you then you define this as the test
34:39
coverage, OK what am I testing guests, we said diversity then
34:42
what is that diversity test case look like?
34:45
Are we actually validating diversity in both ends, right,
34:47
Are we diversity and factfulness should come together.
34:50
So that's the test cars breaking your data into test cases.
34:53
These are test case scenarios. And then right here, just.
34:56
Hold on a SEC. Cause diversity and factual Ness
34:59
come together. Those are two different data
35:02
points, right? Correct.
35:03
What are other data points that would need to come together like
35:07
they're. You gave two data points, but
35:09
potentially what's another scenario where you have two
35:11
different data points that you want to play nicely together
35:14
that cover the test case? So it the factual.
35:19
This usually comes with a lot of stuff because you're trying to
35:22
mitigate the risk, but you don't want to compromise factual
35:26
correctness in your answers, so it comes in in different
35:29
applications, like try to state something up.
35:32
Stuff in my head yeah. If you want to get twitch
35:35
example right, OK, who is that top scorer in in this game
35:39
right. It should it.
35:40
This is an existing just go pose that from from there.
35:43
Now you don't need to. This is factual is it has to be
35:46
correct. You can't optimise for OK, I'm
35:48
going to say the top scorers. Therefore I have to put 2 male
35:52
to female for example right. So now in factual less and this
35:55
is how you design your best, that factfulness should proceed
36:00
in most cases until I know something else could proceed,
36:06
other risks like in this one, but should proceed the risk that
36:09
ideally you ideally proceeds factually right?
36:12
So that now we start thinking about OK, how the come together
36:15
and the tooling can make this make sense of the data
36:19
scientists should not be thinking about what are the gold
36:21
standards out there, What are the evaluation criteria.
36:24
I just need to share my problem and figure my functional
36:28
requirements and I see something populated based on my data.
36:33
How are we going to test for the risks?
36:35
How are we going to verify that my tests for the functional
36:37
requirements and then after that we have the metrics and then the
36:40
acceptable score. Because acceptable score if
36:43
Gemini for example was not a public thing.
36:46
Maybe it's a fun tool that is just meant for to model how
36:49
initially got started the Twitter tool, it just meant to
36:54
be Azure then that could be an acceptable.
36:57
You have high tolerance for this.
36:58
It's fine you put the founding fathers with different
37:01
ethnicity. So again the acceptable
37:04
standards here is the last thing.
37:06
So once you have the test coverage, how you start having
37:09
conversation inside the organisation and this is where
37:11
they gonna send us should look like ask you to to mention that
37:15
you have to put quality test coverage.
37:18
How are you testing your application on your data now?
37:23
These senders or these test cases are left for the team, so
37:27
they find. And I've heard a lot of talk
37:31
about model cards. What's your opinion on those?
37:35
Is it kind of in the same vein as that?
37:38
Yeah, it is. It is an attempt from the
37:40
industry to cover this gap, to close this gap with every model
37:45
provider, whether they're not really talking about Genii,
37:48
don't even, you know networks well provides these share,
37:51
whether it's a marketing effort or like a technical effort or
37:55
academic, they share how my product is performed, right.
37:58
And that's the point of exactly because and that proves the
38:01
point that proves that they need it as providers and their
38:05
customers who was who are the other data scientists building
38:08
applications however they needed.
38:09
So we need to know from building a model and building a new
38:14
company that's gonna compete with Openi.
38:16
Then I want to know how do you think it's performing guest
38:20
opening act Now the missing part that we saw from the industry
38:24
here that it doesn't have the enough trust like there's
38:27
sufficient trust to start work on them from the data sets that
38:30
are being used locally and what they are from the methodology
38:33
that's being used, a clear what it is from the metrics and it
38:36
was acceptable. So all these details and
38:38
obviously from the rest everybody's thinking about risks
38:41
from their own angle and they can put any results in their
38:45
model cars, others. So we want to standardise that
38:47
which is the same pillars that we mentioned before that test
38:51
cases are the coverage, the metrics, the methodology and
38:54
acceptable guidelines are acceptable thresholds.
38:56
Now once we set these as standards, then everybody should
39:01
come in provider them opening. If it's coming out, I will
39:05
benchmark it and I will I will create a model card that follows
39:09
these gold standard principles. What I really like here is that
39:15
once you have the test coverage out there and you understand
39:17
what how it's performing, you don't need only technical people
39:22
in the room. You can have other people.
39:24
And that's the beauty of it, right?
39:25
That's where the subject matter experts can come through or a
39:30
diverse range of selected people can come through and look at the
39:35
results and say, OK, have we thought about this as a test
39:40
coverage? Kate Yes, that's music to my
39:43
ears. This is.
39:43
And then this is where you collectively start chasing that
39:46
long tail of edge cases and in a systematic way like you the long
39:50
Tillbridge kisses and non saying it's now it's just exponentially
39:54
growing in risk with the Genii capabilities.
39:57
Then the chasing that long tail is not something that you do in
40:02
a notebook or something or usually it's in that data
40:04
scientists said OK we've seen these cases before but now you
40:08
have a systematic query and we've seen customers start with
40:10
OK 1020 test cases. That's our understanding of our
40:13
domain today and then few months into.
40:18
Testing with this approach they have hundreds of these test
40:20
cases and now becomes OK now I have I have more understanding
40:25
of my own domain as a company. I collectively and I love what
40:29
you said here about like OK everybody, the product manager
40:31
who is customer comes in, even the customers collaborate and
40:35
leadership because when we start thinking about testing that is
40:38
the name we chose for our company Kolena.
40:40
It's an Egyptian slang for all of us.
40:43
And that's the point of it's like, OK, all of us are
40:45
collectively working together, collaborating in into chasing
40:49
that long tail. Dude, so you wanna know?
40:53
Speaking of Egyptian slaying, my favourite thing about the whole
40:58
conference is that it made people always say the Mohammed
41:05
is the most common name on earth.
41:08
I was like, I don't know that many Muhammads.
41:11
At this conference, we've got three different modes talking.
41:14
And. True to form, we are perfect
41:19
distribution of. This is.
41:22
This reflects actually the real sample here of the of.
41:24
The real sample. Population in the world, yeah,
41:27
it's obviously it's just a coincidence that happened.
41:30
But yeah, this is something that we, especially if you look at
41:34
the yeah, yeah, the name is very common, right.
41:38
And then we, we start seeing here like OK, we need the
41:41
representative from this organisation, we need the the
41:43
leader in this organisation and it starts happening.
41:45
My name is Mohammed too. So it's like, but it's fun.
41:48
It's fun to have everybody there.
41:51
Yeah, you know, I I know I've been kind of a pain in the ass
41:55
the last couple months on trying to make sure that we have
42:00
representation from diverse fields and it is so cool to see
42:06
this coming together. Right now we've got so many
42:09
incredible speakers and I'm so excited because we're, I think
42:14
we're almost hitting the goal that I set out with where I
42:17
wanted to have 50% male, 50% female.
42:20
I don't know if we fully hit it, but we're close, man.
42:24
And that's really cool for me to see because I know that's really
42:29
important. And having two daughters, you
42:31
also have two daughters too, right?
42:33
So. Like a daughter and son, but
42:34
yes. Ohh exactly.
42:36
There we go. Yeah, you have eight daughter
42:40
having two daughters, you have a daughter.
42:42
It makes we're like trying to create the world so that
42:45
hopefully. Absolutely.
42:46
If they never want to get into this field, they don't feel that
42:50
stigmatism or they can see others in the field doing what
42:54
they're doing. Yeah.
42:55
And we talk about representation matters, right?
42:57
Like representation matters. Obviously it matters even more
43:00
when you're trying to standardise the most important
43:03
technology, technology. Down.
43:07
And it's a bunch of men between 25 and 45 talking about white
43:13
males, talking about how we need more diversity, more sensitive.
43:17
A new gold standard. Yeah, no, we're not gonna do.
43:20
That diversity here comes not just in the gender, gender,
43:23
race, yeah, location, domain. We focused a lot, you know,
43:27
that's very well. We're focused on domain being
43:29
represented. We focus on everybody coming is
43:32
a leader and influencer in their own organisations so that we can
43:36
take actual actions with this. So everybody's coming, coming,
43:40
representing their own organisation into this AI
43:42
quality movement. It's going to be very exciting.
43:44
I'm very excited about. And of course we're gonna throw
43:48
in a few of these curveballs from my side where we are going
43:52
to have a little jam room. I think Joe Reese is going to be
43:55
there and he's gonna be able to DJ.
43:58
And we've got my buddy Michael Eric who did a stand up set at
44:03
the last Virtual Conference and he's going to be making fun of
44:07
me and do a little stand up set here too.
44:10
So yeah, I'm super excited man. It's coming together.
44:13
It is going to be a blast. Yeah.
44:16
Excited for this? Yeah, get your tickets early,
44:18
then we're about to sold out probably by in the next few
44:20
weeks. Yep.
44:22
And and of course, like you said, there are tangible
44:26
outcomes that we're shooting for.
44:28
We've got working groups coming out of this, which is awesome to
44:31
see like these special interest groups, the working groups that
44:35
anyone can join. So even if you are not able to
44:38
make it in person, hit us up about the working groups because
44:42
that's how we're going to move the ball forward in the
44:46
industry. Yes, Sir.
44:48
Let's do it. Alright man.
44:51
Well it was great talking to you and I'll see you soon enough.
44:55
We we sink like every other day considering we're yeah we're
44:59
getting speakers we're doing all kinds of stuff for the
45:02
conference. So yeah, but it it's been good
45:04
to have you here. Scott's point Yeah, let's let's
45:08
put this together and want to hear from everybody who's
45:10
watching this episode as we as we built for it.
45:12
We have two months until the conference.
45:13
Let us know your thoughts. Let us know if you're thinking
45:15
about how we can collaborate. If you can't make it in person.
45:18
Let us know if you can make it in person.
45:22
It's just and I know we can meet now when we start thinking this.
45:25
This is a a huge effort that is that requires a lot of people in
45:30
the industry so that we can add that to the EU Act and the nest
45:34
regulations. We can provide something that is
45:36
applicable and is actually something that provides safe.
45:39
I So let's work together as an industry for this.
45:43
It's inspirational. I don't do.
45:45
It. Ohh there we go.