Latest / Agentic Conversations (formally mlops.community) / Reliable LLM Products, Fueled by Feedback // Chinar Movsisyan // #251

Transcript

Reliable LLM Products, Fueled by Feedback // Chinar Movsisyan // #251 · Agentic Conversations (formally mlops.community)

0:00
My name, Chinar Moses, my title I'm a Co founder and the CEO
0:08
here at Feedback Intelligence. We call it feedback intelligence
0:12
and how I take my coffee. I just take it.
0:16
How are they? They give it to me.
0:20
Welcome to the ML OPS Community podcast.
0:22
What is going on everyone? I'm your host, Dimitri OS, and
0:26
today talking with Chinar, we had a conversation about metrics
0:31
and evaluating your AI products, not evaluation metrics.
0:36
She brought a whole new approach and idea that I had not heard
0:39
before and really looking at AI products just like we look at
0:44
other products and figuring out what are the ways that we can
0:49
know if our AI products are successful.
0:52
But not thinking about it from the lens of an engineer needs to
0:58
be the one that is figuring out if the product is successful.
1:02
There are so many other stakeholders in the loop.
1:06
We need to be thinking about how they are able to look at metrics
1:11
around product usage, specifically the PM.
1:15
How does a product manager know if a product is well received or
1:22
not if these AI products, And this comes back to something
1:25
that I have been thinking about frequently when it comes to the
1:30
democratisation of AI. What we've been doing with AI is
1:37
allowing everyone to leverage its power except for the most
1:42
part, the majority of the tools are built for engineers.
1:47
We aren't building as much tooling for the non-technical
1:52
stakeholders who are also in the room as these AI products are
1:55
being shipped. Chinar talks a lot about how we
1:59
can think about the other people that are in the room and
2:03
encourage them to be there. I love this metaphor that she
2:08
used or analogy that she used, which was where is the mix panel
2:14
or amplitude for AI product usage.
2:20
We have not seen that yet. I'm excited for the idea.
2:25
Would love to hear what you all think.
2:26
So drop in a comment. Let me know if this resonates
2:30
with you and as always, if you liked it and especially on this
2:35
one, if you know a product manager that would like this
2:38
kind of stuff, send it on over to him.
2:41
I think product managers everywhere are thinking deeply
2:46
about this and wondering how the hell they're supposed to do it
2:51
with these AI products. But that's my assumption, and
2:54
I'd love to get a little validation on it.
2:57
Let's jump into the conversation with you now.
3:05
I'm managing to stay hydrated on a hot day in Germany, which is
3:10
like one out of three that we have per year.
3:14
So I know you're in San Francisco still and you have to
3:19
deal with the heat a little bit more often.
3:22
But I want to talk about all the different cool stuff that you
3:26
have done up until now. So tell me more about this
3:31
object detection model, putting them onto drones in the
3:35
agriculture industry, Raspberry Pi deployment.
3:38
You're going from computer vision models.
3:41
I imagine they were very big. You had to make them small so
3:44
that they worked on a drone. Give me more context because
3:48
that sounds like an awesome problem to be solving.
3:51
Yeah, yeah, that that was very fun.
3:54
So and it was my first project, to be honest, like for deep
3:58
learning, I did some projects before that, but that was
4:02
something that I was, oh, this is going to be not a research
4:06
project, but something that is going to be in production.
4:10
So I should be careful. We should be careful.
4:13
And I remember that was like Yolo 3.
4:17
Like that was that time we were yeah, we were excited about
4:21
that. And in terms of draws, like
4:25
solving the problem for object detection, it's all about high
4:29
resolution images. So very like tiny objects like
4:34
you, you have to like for crop, crop analysis or a lot of big
4:40
this stuff or spraying like automatically spray the the
4:46
field. So this was the project and
4:49
where we had a lot of data, like more than more than 10K high
4:53
resolution images. And we, we, we were required to
4:58
annotate it and we had a lot of people there to manually go over
5:02
it and then train it like annotate the data and then fine
5:07
tune that Yolo 3 adapted to that use use case like custom
5:12
training and then get some models, some checkpoints.
5:16
And of course, it's not going to be to be good enough on, on
5:22
Raspberry Pi because of JAWS like you cannot have GPUs on, on
5:27
JAWS, unfortunately. And the idea was just translate
5:31
it to lighter model and deploy it and then and then see what's
5:36
happening. Even though after doing this, we
5:39
got a lot of like bugs we had, we had a lot of like false
5:44
positives, false negatives because of elimination lighting,
5:47
these old stuff, different things were happening.
5:52
But yeah. I'm still waiting for the
5:54
company that straps a GPU to a drone.
5:58
Maybe NVIDIA? Yeah, that's the next billion
6:02
dollar company, huh? That is who wants to go and
6:06
raise some money for that one? That is hilarious to think
6:09
about. But the, the interesting piece
6:11
there, I think is you were doing this with Yellow 3.
6:15
As you mentioned, it was deep learning before deep learning
6:20
was cool. Most people I imagine were
6:25
excited to work on deep learning back in those days because
6:29
everybody learns about it. It's the cutting edge
6:31
techniques. And I can only echo what I've
6:36
heard from the community and all these conversations and how
6:41
everybody wanted to work on deep learning until they got into the
6:44
enterprise. And then it's like, oh, yeah,
6:47
like, there's not really deep learning happening here.
6:49
I have to go and do that. You didn't have that case, and
6:53
you went out and also did other cool stuff with drones, right?
6:56
Like, that wasn't your first foray into it.
6:58
Later you were doing surveillance stuff and working
7:01
with yellow models. Yeah, yeah, it was.
7:04
It was a surveillance use case for person and car detection for
7:10
street surveillance purposes. And it was a small project, like
7:14
a small like startup. We can say at that time I was in
7:19
Armenia, based in Armenia and we were doing that for like year
7:23
one. And but the, the idea was very
7:26
cool. Like how we can just have it
7:28
like person detection, car detection and deployed in front
7:31
of like shops, supermarkets and just have that for not manually
7:36
count. What is the like monitor the,
7:39
the camera data, like 24 hours footage, but just have some AI
7:44
and automatically like have some basic analytics.
7:47
What's happening outside? Yeah, you don't have to have
7:50
somebody with a clicker clicking every time a car drives by,
7:55
which makes everyone's life a lot easier, I can imagine, and
7:58
frees up some time from people that The other piece that I
8:02
wanted to talk to you about was, was the whole like generative AI
8:07
stuff you were doing before generative AI?
8:09
But before we do that, did you come up with any cool techniques
8:15
or ways to get these big computer vision models smaller?
8:20
Were you pruning in those days? Were you distilling them?
8:23
What was it looking like? I feel like that's one of the
8:28
one of the one of the challenges that's right now still people
8:34
are facing because these all huge foundation models or
8:37
Transformers are like if we are able to deploy it on, they still
8:43
eat and deploy it on small like where there is no GPU stuff,
8:48
there is no computational erasers available.
8:53
So that's the that's the solution there.
8:56
So at that time we were doing just a translation from Python
9:00
to C++, which was very helpful. With that.
9:03
We were able to distil it and we were able to keep the accuracy.
9:08
And on the other note, keep the, the speed like how many frames
9:13
we can process per second, which is very important for JAWS where
9:17
you just get a lot of data, a lot of video data and you need
9:20
to process it and you cannot just skip, skip frames there.
9:26
So this was the the technique that we came up and it was good,
9:31
like good enough for, for that use case.
9:33
But I imagine that for a lot of use cases, especially like draws
9:38
are being used in military or other like verticals.
9:43
For sure there should be some other techniques to have banners
9:47
as accurate as possible, like close to 100%.
9:50
OK, 20 seconds before we jump back into the show, we've got
9:53
ACFP out right now for the Data Engineering for AI and ML
9:59
virtual conference that's coming up on September 12th.
10:02
If you think you've got something interesting to say
10:05
around any of these topics, we would love to hear from you.
10:09
Hit that link in the description and fill out the CFP.
10:12
Some interesting topics that you might want to touch on could be
10:17
ingestion, storage or analysis like data warehouse, reverse
10:21
ETLSDBT techniques, etcetera, etcetera.
10:23
Data for inference or training, AKA feature platforms if you're
10:27
using them, how you using them, all that fun stuff.
10:29
Data for ML, observability and anything fin OPS that has to do
10:34
with the data platform. I love hearing about that.
10:36
How you saving money? How you making money with your
10:39
data? Let's get back into the show.
10:42
So now tell me about this plastic surgery stuff that you
10:45
were doing. So I got that project from
10:50
doctors like they were doing plastic surgery and their
10:55
question was their challenge was, hey, we are using Photoshop
10:59
very like classic and we have a lot of data like images of faces
11:06
and we need to like set the landmarks, understand maybe nose
11:11
should be shaped this way, etcetera.
11:13
Like a lot of manual work and then back and forth with the
11:17
with the passion. So whether he or she likes the
11:21
the new shape or not. So this is the question and
11:24
we're like, OK, how we can solve this?
11:27
And we take we took that those all kind of like they had a lot
11:32
of data reference like reference images, like pairs where they
11:39
already manually did that And we trained a peaks to peaks model.
11:44
I remember it was like by Taser flow and it was con based like
11:49
generative adversal network based solution.
11:52
And then after that, doctors were able to upload their like
11:57
images, like the actual image, a patient and then get generated
12:04
image with a new face, like new nose or something else that
12:10
should be. It was very fun project,
12:12
especially having those old like images, like a lot of like these
12:17
different faces, different shapes.
12:20
And then trying to have these generative adversary networks
12:23
trained in a way to have better good results for for other like
12:29
nuts the data not in the training site.
12:33
The one of the first people that I talked to when diffusion
12:37
models were getting really popular was a friend of mine
12:40
who's a dentist. And I was showing him some of
12:43
the cool stuff you could do with, at that time, Stable
12:46
Diffusion. And he instantly was like, huh,
12:49
I wonder if I could do this with people's teeth.
12:52
You know, I have to constantly be thinking about how to better
12:57
make the mouth and the teeth and the shapes of the teeth,
13:01
especially if he's doing work on them or he's putting in new
13:05
teeth. And, and I was like, I bet you
13:09
could, but I think there's probably a better way to do it.
13:11
Like, I would be a little bit worried of for a diffusion
13:15
model. Yeah.
13:16
To give someone a new teeth or a new tooth.
13:19
Yeah, yeah, yeah. That, that, that's very funny.
13:22
And also I feel like that's helpful for not technical for
13:25
doctors having that like, yeah, some applications deployed on
13:31
their computer just in case to have some reference before doing
13:35
their manual work. For sure they have to check it.
13:38
You don't know what what kind of generated image could be there.
13:43
For sure there are false positives of both negatives.
13:45
We cannot guarantee 100%. But yeah, that was really fun
13:49
and also like good project. Yeah, it gives them a little
13:53
inspiration, so hopefully that helps.
13:56
Now I for sure want to talk about like the idea of basically
14:05
highly regulated industries and you've, it feels like you've
14:10
worked in some different highly regulated industries like
14:14
healthcare. Are there gotchas that you've
14:16
seen that keep models or AI from making it to production?
14:24
One of the patterns that I can mention here is in these all
14:31
verticals like from agriculture industry to healthcare
14:37
surveillance. Now like fintech, we are
14:40
involved in, I'm involved in Fintech now, but there is this
14:44
like gap between domain experts and users.
14:49
And as like as a builder, that was something that we always
14:54
deal with. And now as well, like we always
14:57
think about building something. But they're the model actual
15:01
users are people who who think that this is a magic like, oh,
15:06
we need to get this out of like doctors for plastic surgery
15:11
purposes. We need to get this output.
15:13
But like, there are a lot of cases where it's hard to
15:17
communicate that output, why that that is happening and why
15:23
you should expect that as well. Yeah.
15:28
Have you found any strategies on how to better communicate that?
15:32
No, I, I, I I I cannot say I found.
15:37
I remember when I I was doing my PhD and one like I had I had my
15:42
supervisor, but also I had a doctor, professor, supervisor.
15:45
The first two months were horrible just to kinda have some
15:50
language to talk about. I was saying something and she
15:54
was like she's an she was, she is an expert, I mean professor
15:59
in cardiovascular field. But it was very hard to find
16:03
this ground truth and talk to each other just to understand,
16:10
even though we were solving the same problem.
16:12
And so you think that's because you didn't have that shared
16:15
vocabulary? Or was it just because the
16:18
understanding or the expectations from the
16:21
non-technical side were way out of scope of what is actually
16:26
possible? I was saying both but the second
16:29
one later one is it has more weight than the first one.
16:35
Like this all expectation and also like understanding that the
16:41
expectations and understanding what's what's happening with AI.
16:46
What is that like? What is the definition of AI?
16:49
Yeah, yeah, I feel strongly about this.
16:53
And I feel like these days, especially because there is so
16:57
much noise around AI, it's very hard to get non-technical
17:02
stakeholders and even technical stakeholders.
17:04
Like there's so many technical people that I've talked to who
17:07
have played around with LLMS and they still think that it will do
17:12
way more than it is actually capable of doing.
17:14
And so being able to properly manage those expectations is an
17:21
absolute art. Yeah, yeah, yeah, 100% and and I
17:26
feel like that will help a lot in terms of product like any AI
17:32
based product development and and making it to production.
17:39
That's the goal, right. We are not doing any more well.
17:41
We are doing research, but the goal is to have it in
17:44
production. Exactly.
17:47
And so Speaking of these almost like product metrics and jumping
17:54
to that area, I think there is some fun stuff that we can look
17:58
at when it comes to how you like to measure product metrics, how
18:04
you like to think about evaluation metrics.
18:07
Almost like even we can go into monitoring and observability and
18:12
what you've been seeing out there.
18:14
Because I think you have a different viewpoint than most
18:18
people that I've talked to where we know that evaluation is the
18:22
hottest topic right now. We know that everyone is talking
18:25
about evaluation, but it's almost like nobody that has
18:29
really cracked that nut and figured out the tried and true
18:33
and tested way to evaluate output.
18:37
So what's your take on this whole scenario?
18:41
Oh, this is one of my favourite topics in terms of evaluation of
18:46
super Realty monitoring and then this like product analytics,
18:51
diagnosis, understanding like AI based product, no classic web
18:55
applications. So like in this topic, one of
19:00
the examples that I try to bring, I try to bring as an
19:06
analogy is that is this whenever we build any web application and
19:11
like just host it, what do we do?
19:13
We use some analytics, right? We are not waiting for a couple
19:17
of months and then turning on this Google Analytics or
19:22
Amplitude after six months. No, whenever we host it, we just
19:26
link it and then every day we check it, right?
19:28
We check what's happening. They use this button, they don't
19:31
use this. So my question is why we are not
19:34
doing the same thing with AI based solutions, right?
19:38
We only care about evaluation, which is way important.
19:43
That's important. It's not a, I'm not questioning
19:45
that. Yeah, we should evaluate it.
19:48
We should like calculate these old metrics like if one score or
19:53
like accuracy, recall, map all these metrics.
19:57
Those are very important. But once it goes to production,
20:01
we should have something like some data-driven way of
20:04
understanding well how it is being used, what is, how end
20:09
users are using it, what are their expectations, what they
20:13
want actually from this product. And then catch it like fetch it
20:18
and then like do this cycle again and again.
20:22
Like I feel like we are taking from model to user direction.
20:27
But right now as we all are thinking about production LLM
20:33
based solutions, maybe we need to take the other direction like
20:37
backwards from users to model. So this is something that I feel
20:42
like we are missing in this like, well, AI evaluation,
20:49
observability, monitoring observability like chain.
20:54
So how do you see that playing out in practise?
20:56
Is it the conversations like, so let's just take the
21:01
quintessential Rag chatbot. It is the conversations you're
21:07
having with the chatbot or is it beyond that?
21:12
And I know that a lot of people will put the thumbs up, thumbs
21:16
down, which is a pretty bad metric.
21:20
Honestly. I think we've all recognised
21:23
that doesn't really work because it's unclear if it's thumbs
21:27
down, is it? Why is it thumbs down And what
21:30
does that actually mean? I am not the biggest fan of
21:35
chatbots, but let's just like stick to that example.
21:37
How do you look at other metrics around it besides when somebody
21:42
is getting very angry at the chat bot?
21:45
Yeah, yeah. So basically we can't say that
21:48
there are two options here in terms of feedback, in terms of
21:53
understanding how the chat bot is being used.
21:57
One of the options is explicit feedback, which is thumbs up,
22:02
thumbs down, or a plain text about.
22:05
Well, I'm not happy with the output.
22:08
This is wrong. I I I did a query to get
22:11
customer transactions in Q 4/20/22, but I'm getting
22:17
business transactions. This is not correct.
22:19
Imagine financial analysis using this chat bot.
22:22
This is like explicit feedback. But what about all implicit like
22:28
feedback that we can derive? We can get it and then derive
22:33
signals and we we saw that. What I have seen based on my
22:39
experience, just the usage of chatbot like querying, querying
22:45
something, paraphrasing it, leaving the chatbot or causing
22:50
it and then opening it again. These all signals can make a
22:55
difference in terms of understanding what is the issue.
22:59
Maybe there is an issue in terms of knowledge hole in rack system
23:04
like context is not informative enough to retrieve business
23:09
transaction or vice versa. Well, it also feels like, all
23:17
right, maybe some kind of document is retrieved or you
23:21
give a link to where you grabbed a snippet of that document and
23:25
then that link is clicked upon. And so these are all signals.
23:31
How do you make sense of which signals are good and which are
23:34
bad? Because I'm, I'm thinking about
23:35
this and I'm like, is that good? If they click out of the chat
23:38
bot, does that mean that the answer, they've gotten their
23:41
answer? Or does that mean that they're
23:42
frustrated and they don't want it anymore?
23:45
That's a very. That's one of the hardest
23:47
problems. Like I frame it as a root cause
23:50
analysis. So root cause analysis in
23:53
general, it's a like term that we know we have known for a
23:56
while, like for decades, right, for traditional applications and
24:01
now root cause analysis for LLM based solutions, LLM based
24:06
products. So in order to understand those
24:09
signals like negative, positive, neutral just having these like
24:13
logs like queries, responses and then some information about
24:20
knowledge like context maybe retrieval like system is
24:23
available like chain size, all these things.
24:26
How we can use all these like inputs and perform some root
24:32
cause analysis and derive these all signals and then based on
24:37
those signals maybe get some list of actions to automate this
24:43
whole process? So it's basically trying to
24:46
create metadata on all of the information that's going like
24:51
the chunk size or which chunks were fed into the LLM.
24:55
Are you also looking at which embedding model was used to
24:58
create the embeddings? Are you looking at that far
25:00
back? How do you?
25:02
So far, no, we haven't done that that part yet.
25:07
But we are I, I am I, I, I bring this analogy like onion layers
25:14
like you can do this more and more, but it depends on the data
25:18
that you have access to. If you have access to this like
25:23
only the hyper parameters of retrieval system and and a
25:28
couple of other hyper parameters like temperature, yeah, you can
25:32
do some analysis based on that and provide a list of actions
25:38
about that hyper parameters as well.
25:41
And when you have access to embedded model, yeah, you can do
25:45
more and more, but it depends on what information you have access
25:50
to. So yeah.
25:52
Yeah, that that makes a lot of sense.
25:54
And basically you're saying, all right, let's get all the
25:58
analytics on what the situation is instead of just the actual
26:05
text that the user wrote back or the output text.
26:10
It's like trying to capture every relevant piece of data so
26:14
you can fully analyse the output or the experience that someone
26:20
had with that chat bot. As opposed to that one question,
26:24
did it get answered? What would the correct answer be
26:27
versus what the output of the LLM is?
26:30
You're like, how can we get a full picture of what's going on
26:33
here? Yes, yes, yeah.
26:37
OK, So what are other? So that's a chat bot.
26:39
What are other ways that you've seen this happening?
26:43
Yeah, like chat bots are one of the one of the interfaces that's
26:48
right now LLM based solutions are deployed right.
26:53
Other besides chat bots, we have like a genetic work flows or or
26:59
like conversational AI, just a dashboard like analytics for
27:06
intelligent document processing where again, they can do some
27:09
queries and based on that, those queries they see some analytics
27:15
or yeah, agentic workflows are going to be the next like big
27:20
thing they are. They are being deployed in a lot
27:23
of verticals and root cause analysis of agents, agentic
27:27
workflows is very like. It's more more difficult than
27:34
just root cause analysis of of like RAG or prompt engineering.
27:41
Yeah, break that down for me. What are things that we need to
27:44
be aware of as we're trying to root cause the agentic workflow?
27:49
Oh, in order to root cause, root cause, this agentic workflow,
27:57
agenting workflow like the the implementation was being done
28:02
based on LLMS RAG like small different like ways of
28:10
implementation of LLMS, right Prompt engineering these old
28:14
tests. So now for a genetic workflow,
28:18
it's like orchestration of other root cause analysis, like root
28:22
cause one, root cause 2 and then orchestration of this.
28:26
So that's why I'm saying it's more difficult than the the like
28:32
classic like rag we can call it. But we should solve this to be
28:38
able to solve the orchestrated problem.
28:43
Yes. So it's not only that basically
28:46
you're just going up in complexity, you're multiplying
28:49
the complexity every time you add a new agent.
28:53
And So what are some other ways that you are looking at the data
28:58
to be able to root cause it? Like I I imagine you can get as
29:03
creative as possible, but have you seen big levers that give
29:10
you a much clearer picture? Can you elaborate on the
29:15
question a bit more? Yeah.
29:17
So, so I could see myself thinking, wow, if I can get all
29:25
these different metrics like the all the metadata around it and
29:32
what browser the person is using or what all, everything you, you
29:36
can get very granular with the data that you're trying to
29:39
capture to give yourself that full story and that big picture.
29:43
But at the end of the day, you may find that there are three or
29:47
four metrics that give you 80% of that story.
29:51
Yeah, Yeah. That's a, that's a really good
29:53
question. So we can say we can frame it.
29:56
What what is that base? What is that base combination
30:02
that gives you that holistic view of your product of your
30:07
generative AI based solution? Yeah, that's a that's a good
30:10
one. I don't have an answer, but
30:14
basically it's all about like peeling those old layers,
30:20
identifying these old metrics and then understanding like
30:25
which we metrics can be used as a can can present the other like
30:32
3 metrics over there. Like maybe you can use the one
30:37
and just have the picture of your product.
30:41
And how do we know like is there a clear signals that you've seen
30:45
where you're like, oh, when you find those metrics and it's
30:51
probably different for every use case, but when you do find those
30:55
metrics, you recognise them because they are so different or
31:02
they are such an anomaly? Or is it like you're, you're
31:05
really like just going off of intuition and instinct?
31:10
You know what we did recently? We introduced this impact scores
31:15
idea and this impact score is being calculated using different
31:22
metrics, different like several metrics, the result of several
31:27
metrics. And with this we can kind of
31:30
have an average and we were able to kind of introduce one like
31:38
metrics numbers. So they can they can use that
31:42
and they they can try to make some decision based on that
31:47
impact score. And right now we are iterating
31:49
over that and trying to understand from fintech to legal
31:53
AI how it is like it is different.
31:58
Oh yeah. And the impact score, if I'm
32:01
understanding that correctly, it's just the metrics like how
32:06
impact, how impactful this metric is to the greater picture
32:12
or the the big metric that you're trying to move the needle
32:15
on. So like how are these small
32:17
metrics moving the bigger metric?
32:18
So impact score, imagine like there are issues and this issue
32:25
you can, you can present, you can, you can describe using a
32:33
couple of metrics, but which one has more weight in terms of in
32:41
terms of analysing it farther. So there is no one ground truth,
32:47
right? Which one we should take into
32:49
account Like for issue one, recall is X, for issue 2 recall
32:56
is Y. And now how we can compare it?
33:01
So we kind of map it to 1 dimension and then we do
33:08
comparison. So issues are being compared
33:13
only using impact score. What is the impact score of this
33:17
issue across different customers, across different
33:21
models? So how repetitive is that issue?
33:26
Oh, so I think I'm seeing that you're saying, you're saying
33:32
like, hey, this metric may be very high for some use cases or
33:36
some instances, maybe not use cases.
33:39
It's more like the instance, the interaction with this end user,
33:43
this metric is off the charts. But then on other instances,
33:47
this metric is pretty much non existent.
33:49
So why is that? How impactful is this metric?
33:53
Is it a little bit wonky? What does it really tell the
33:56
story of? Yeah, yeah, yeah.
33:59
What I'm seeing is that we need data analysts for our AI metrics
34:09
and our AI usage metrics really. Yeah, yeah, yeah.
34:12
And we see that it's not only about data analysts, but also
34:18
product managers. These people are involved in the
34:22
process like they are in charge of like in charge of product
34:27
usage, right? It's their role, but also it's
34:32
not a classic like it's like they're this is a bit new like
34:37
to understand these all metrics and and they're like root cause
34:43
of those like numbers. Yeah.
34:46
It's like you have to have a bit of the understanding of what a
34:50
data analyst does or how a data analyst thinks.
34:55
You have to be able to get intimate with that data, but
34:59
it's not necessarily looking at how a user is interacting with a
35:07
product, a traditional product, or how a how someone is buying
35:14
something from an e-commerce store.
35:15
Like it's not that type of data analyst or the, the internal
35:19
data that a company has and analysing how to boost revenue
35:23
or whatever it may be. It's, it's very much, and I've
35:26
heard this from, I can't remember who right now, my brain
35:29
is failing me, but they were saying how data scientists are
35:36
very much in a great position because exactly what you're
35:40
saying sounds a lot like what a data scientist could excel at.
35:45
Yeah, Yeah. So what also I, I noticed with
35:50
this AI like adaptation LLM, let's say let's call it LLM
35:56
generative AI or whatever is like eval tools evaluation like
36:04
environment for developers. But these product managers are
36:08
you between and they need something to know like about
36:12
these products, right? Well, they cannot use product
36:16
managers. They can, I don't think they can
36:18
use eval, eval tools like open source git clothes, old things
36:22
like, but they need something like mix panel a bit more or
36:30
Google Analytics a bit more with like more features like for data
36:36
analysis. So kind of we need a combination
36:39
of of these all like existing things into one place so
36:47
different stakeholders can can do their job.
36:52
And now that is a really key point because you're talking
36:56
about how the product manager needs to understand the value of
37:01
the AI product. And just using the evaluation
37:07
tooling that is built for software engineers and machine
37:11
learning engineers or AI engineers, that doesn't serve
37:15
them very well. No, yeah.
37:19
It's that brings up another great point that, and it's
37:22
almost like the same idea, but because now we have so many
37:30
different personas that are able to use AI and the traditional
37:35
tooling like the infrastructure tooling and tooling in general
37:39
is built for engineers. Kind of like I'm just making a
37:43
broad sweeping general statement.
37:46
The majority of the tooling is built for engineers.
37:48
You get the marketer who has been using low code, no code
37:53
tools and they've been able to create some kind of A use case
37:58
that generates a lot of cash, but they have no way to track
38:01
their prompts, right? And it's like they're not going
38:04
to whip up weights and biases or ML flow.
38:07
No, that's not built for them. And so there's these tools that
38:12
will need to come on to the market that are specifically for
38:16
the low code, no code users or the people that aren't
38:20
interested in the software tooling side of things.
38:24
But they still need to do their job with the AI products that
38:29
are going out. Yeah.
38:30
Yeah, yeah. And on the other note, that
38:34
tooling should be something because what we have seen, what
38:39
I have seen with this product managers and everything they are
38:42
seeing that we try to kind of communicate these all things to
38:48
AI engineers like using mirror. We write down things and I'm
38:53
like, how like that's not a good workflow.
38:57
How do they use it like, and they were like, we translate
39:02
that's like product, let's say like financial analysis using
39:06
I'm a product manager. I'm dealing with that financial
39:09
analyst at Bank of America and my AI engineer doesn't want to
39:14
work like doesn't want to deal like of course, no.
39:18
And then I need to translate that to this AI engineer.
39:21
I need to get that information to write down structure it for
39:25
AI engineer maybe using mirror. I have heard that like they are
39:30
using mirror. I'm like how it's possible?
39:32
Like you have a lot of tools like these people have a lot of
39:35
eval tools, a lot of things they also do you use also JIRA?
39:40
Like no, I there should be something like these old people
39:44
can just log in and everyone will be happy, right?
39:47
Well, even just thinking about if you are the product manager
39:52
and you're doing what you were just talking about, let's gather
39:57
as much data as possible and analyse this data so that we can
40:01
get a full story on how effective our AI product is,
40:06
right? Let's let's say that is an
40:09
awesome way of deciding if an AI product is valuable or not.
40:15
How do you put that into practise, right?
40:16
Like is the AI engineer, you're going to call up your data
40:19
engineering friends, You're going to be like, OK, so now all
40:22
of a sudden we've got a tonne of data.
40:24
We got to figure out where we're going to store it, how we're
40:25
going to present it to our product managers, how we're
40:28
going to do all of that. You now are going through that
40:31
cycle and you're doing what you do with traditional products,
40:38
but in a little bit wonky of a way.
40:43
I would say it's, it's almost like, and, and it's so new that
40:47
you're, you're kind of searching in the dark and you're figuring
40:52
things out as you go and you're recognising, oh, do we also need
40:56
that data? Maybe that could be good.
40:58
Whatever. Let's throw it in there just in
41:01
case. Yeah, yeah, exactly.
41:03
Maybe we need that for the for the 2nd iteration.
41:06
And we have heard this as well. Like, you know, I was
41:10
interviewing this small AI team and they were, they are
41:14
delivering to financial like institutions, like finance now
41:21
technical people. And one of the I asked like,
41:23
what is the challenge right now? Is it about drag?
41:26
What is it? What is it about?
41:27
And he was like, it's, it's about under translating their
41:34
like these end users experience in terms of product usage to
41:44
like code requirement to feature requirement.
41:48
This like translation is like what kind of data we need, how
41:53
to process that data, how to understand them and how
41:57
incorporate that information to let's say RAG implementation.
42:02
So like there is a kind of gap, like right between these two
42:07
people, even though they have the same mission.
42:10
Let's have an LLM in production. That is 100% something that I
42:16
can understand and vibe with. It's a, it's like there is a
42:23
disconnect, but it's not a disconnect on purpose.
42:27
It's just a disconnect in the way that things have shaped up.
42:31
It's almost like we've ran really, really fast on the
42:35
engineering side and we've been trying to keep up on all the
42:39
other sides of the business. But since the exciting part is
42:45
all the cool stuff you can do on the engineering side, that's
42:48
where the energy's been going into and that's what people have
42:51
been focusing on. And so I see the, and now I
42:55
understand completely what you're saying in terms of
43:00
evaluation is something that we've been thinking of and we've
43:03
been talking about, but we almost need to look at it from a
43:08
different lens as opposed to only the engineers looking at
43:13
evaluation metrics and evaluating the output.
43:17
We need to bring in more stakeholders.
43:19
Yeah, yeah, definitely. Because this is not a one side
43:24
problem. We look at that they, they say
43:27
like you are a little engineer, you have to deliver 100%
43:31
accurate chat. But but I mean, it's not only my
43:34
problem, right? There are a lot of people
43:37
involved in this process and it's kind of something like they
43:41
get this business problem. They try to kind of explain it
43:46
to me as an engineer and then they are not happy with my work.
43:49
But right, right. Like these people should take
43:53
care of that as well. Like how can I do that alone,
43:56
right? And then have this back and
43:59
forth. So like I have had that before
44:01
this like generative I I remember when I was delivering
44:05
that like culture, I mean that object detection models,
44:11
everything, they were like time to time they were coming back or
44:15
it's not working well on the way they're changed.
44:18
I mean, now it's very rainy this and because of that we, we
44:23
didn't have much data. I didn't know that this this
44:26
fall we will have a lot of rain. So this kind of like, you know,
44:29
but it should be something like everyone is aligned some
44:33
platform toolkit. So they see same is happening
44:36
with LLM based solutions. It goes back to the big
44:40
disconnect that you've seen with the subject matter experts not
44:46
being able to properly work with the engineers and map out all
44:52
these different use cases or all these failure modes or the whole
44:57
user journey, whatever it may be that you're trying to explain.
45:02
You have so many different people that need to be involved
45:05
and we as a community and as like just people working on LLMS
45:15
have almost over optimised it feels like for the engineering
45:20
side of the house and now we need to play catch up on all the
45:24
other sides of the house and how can we get every stakeholder
45:28
properly in the room, having the conversation and making sure
45:33
that the project is successful. Yeah, yeah, 100%.
45:37
And compared to like if I compare my life, the problem
45:43
that I was facing like 6-7 years ago with this like object
45:48
detection task. And now the difference is that
45:54
right now with this generative AILM based solutions like
45:57
chatbot, it is more user centric.
46:01
Like bunch of users are using it.
46:04
But for like computer vision model, it was just like, let's
46:08
say 5 draws with this like spraying model to spray the
46:15
field automatically. Yeah, that was a different like
46:18
very like 5 complaint messages. OK, we will do something.
46:23
But like hard, like dozens of like, I don't know, a lot of
46:28
thousands of users are using Git.
46:31
And then you're getting this all like you don't have any
46:34
insights, you don't have any data-driven way to understand.
46:37
And then there is this disconnection as you call it.
46:41
This is fascinating to really harp on and it's exciting to
46:45
think about how much potential there is because of this,
46:53
basically this gaping hole in the way that we're doing things
46:56
now and almost like a not a utopia, but just, hey, wouldn't
47:00
it be nice if we also had other people that could be dealing
47:04
with this and helping us out and bringing their wisdom and
47:07
expertise into the picture? Yeah, yeah, yeah.
47:13
This is something that should be solved by.
47:17
It's a multi angle problem. It's not only about engineering,
47:22
it's not only about product, it's not only about maybe
47:25
customer success, but it's all about like these all like
47:28
aligning these all stakeholders into one place.
47:31
And and yeah, I feel like with that we will have some some
47:36
approach to solve the problem. 100% All right, before we go, I
47:41
got to know what's your deal with flamingos?
47:44
I'm like, I didn't remember that.
47:49
Oh, I love them. You, you just saw that.
47:52
My hair is pink, so I love them so much.
47:56
The design and everything we have, it's all about pink and
47:58
purple. So it's just a laugh.
48:00
You cannot do anything with that.
48:04
Well, this has been awesome. I really appreciate you coming
48:06
on here and thinking something that I have had a hard time
48:11
articulating, but it now that you say it, it's like, yeah, of
48:15
course. And I really like the fact that
48:17
you've gone out there, you've been talking to a whole lot of
48:21
users that are trying to capitalise on AI.
48:26
They're, they've got AI in production and now it's almost
48:29
like the day two of, OK, So usually with like when we ship
48:36
products, we have observability and like we want to know how
48:40
it's doing, but we don't really have that clarity on it.
48:44
And so coming on here talking about these issues and hopefully
48:49
there are people out there that have been seeing this too and
48:53
they've been recognising it. And I would encourage anyone
48:57
that is also on the same wavelength to reach out to you.
49:01
Awesome. Thank you so much.
49:02
Thank you for having me. This was awesome.