Latest / The DEX Show: A Show for IT Change Makers / 🎙️ Reality Bytes: The AI-Powered Future of DEX w/ Mohamed & Maxime 🔍

Transcript

🎙️ Reality Bytes: The AI-Powered Future of DEX w/ Mohamed & Maxime 🔍 · The DEX Show: A Show for IT Change Makers

• 0:01
  Speaker 1: Welcome it, change Makers to the deck Show with Tim
• 0:05
  Speaker 1: Flower and Tom McGrath. Let's get into it.
• 0:09
  Speaker 2: Do you know what it is, change Makers? It's a
• 0:11
  Speaker 2: deck show show within a show within a show where
• 0:14
  Speaker 2: Tim we get to talk to fascinating next thinkers and
• 0:17
  Speaker 2: I do think you know how many organizations could could
• 0:20
  Speaker 2: vi even.
• 0:21
  Speaker 3: Work with Tim, and I love that it works.
• 0:23
  Speaker 4: We got I love talking with people outside and get
• 0:25
  Speaker 4: thought leaders outside, but we got some really smart people
• 0:28
  Speaker 4: inside and it's great to be able to share it
• 0:29
  Speaker 4: with everybody.
• 0:31
  Speaker 3: It's absolutely so.
• 0:32
  Speaker 2: And you know, before we introduced today's guests, Today's Next Thinkers,
• 0:37
  Speaker 2: because they are plural, we should just mention but that's
• 0:40
  Speaker 2: why you don't hear the usual reality by its panel.
• 0:43
  Speaker 2: We thought we've got two guests. We don't want to
• 0:44
  Speaker 2: have too many people on the microphone be a bit
• 0:47
  Speaker 2: disorientating for us all but today's guests. We invited them
• 0:53
  Speaker 2: on because they actually co wrote a Next Think blog
• 0:56
  Speaker 2: together called Teaching AI to Speak Next Thing Query Language
• 1:00
  Speaker 2: Lessons from Next Think Assist, and we thought it was
• 1:04
  Speaker 2: such a fantastic story we would have them on the
• 1:06
  Speaker 2: show to explore the conversation, to explore the topic rather
• 1:10
  Speaker 2: in a more conversational format, So let's without further ado.
• 1:14
  Speaker 2: First of all, welcome back to the show, returning guest
• 1:18
  Speaker 2: Mohammed CAFSI, Director of AI Diagnostics here at next Think, Hi, Mohammed,
• 1:24
  Speaker 2: And let's also welcome first time DEX Show guest Maxim Kinnitt,
• 1:30
  Speaker 2: software engineer and tech leading diagnostics here at Next Think Hi, Maxim,
• 1:35
  Speaker 2: Welcome to the show for the first time. And you both,
• 1:38
  Speaker 2: of course work on the AI side of the business,
• 1:40
  Speaker 2: and that's very good timing because there's a bit of
• 1:44
  Speaker 2: AI assist related news we should touch upon soon enough
• 1:48
  Speaker 2: in our conversation. But before we get into the technical
• 1:52
  Speaker 2: aspects of assist in general and the story we're here
• 1:57
  Speaker 2: to discuss, just want to be by asking you both
• 2:01
  Speaker 2: a question about your relationship with this technology, if you will.
• 2:07
  Speaker 3: Was there a particular.
• 2:08
  Speaker 2: Encounter with AI that made you think, Wow, this is
• 2:13
  Speaker 2: going to change everything and it could be something for
• 2:15
  Speaker 2: me personal life, you studied life, your professional life, any
• 2:20
  Speaker 2: fear above, Maxim, since your first time guest, maybe you
• 2:24
  Speaker 2: should start for me.
• 2:25
  Speaker 5: Actually, I think it was here in the office. I
• 2:28
  Speaker 5: was working with some colleagues and we were per programming.
• 2:32
  Speaker 5: Actually because I wanted someone to explain me how to
• 2:34
  Speaker 5: do something, and I was sharing my screen and at
• 2:38
  Speaker 5: that time I had just been on board on a
• 2:40
  Speaker 5: co pilot technical preview program and I just enabled it.
• 2:46
  Speaker 5: And while I was starting to program what my colleague
• 2:50
  Speaker 5: was explaining, at some point, Copilot came and generate a
• 2:54
  Speaker 5: huge chunk of codes and we all looked at it,
• 2:58
  Speaker 5: and there was a black in the conversation and someone said, well,
• 3:03
  Speaker 5: that's a hell of a completion, and it was exactly
• 3:05
  Speaker 5: what we needed. And I think this is when I realized, Okay,
• 3:09
  Speaker 5: this is going to impact ue tremendously.
• 3:15
  Speaker 4: For me.
• 3:15
  Speaker 6: This is my moment.
• 3:16
  Speaker 3: I like it. I like it. What about you, Muhammed? Yeah,
• 3:20
  Speaker 3: that's a great example, but mine is much less.
• 3:22
  Speaker 1: A much older example. So's it traces back let's say
• 3:30
  Speaker 1: twenty or twenty two years ago when I started, like
• 3:33
  Speaker 1: my computer science degree here at THEPFL which is the
• 3:37
  Speaker 1: engineering school in Lausanne, and it was like typically my
• 3:42
  Speaker 1: first oncounter with let's say AI, discovering AI. And I
• 3:46
  Speaker 1: remember we had like a you know, a software engineering
• 3:48
  Speaker 1: project where we basically have teams and we're building We
• 3:51
  Speaker 1: had to build a you know, a checker game. Do
• 3:53
  Speaker 1: you know checkers?
• 3:55
  Speaker 3: Sure? Yeah, yeah, And we.
• 3:56
  Speaker 1: Had to build the checker games and every team has
• 3:58
  Speaker 1: basically to code the checker game from scratch. And the
• 4:04
  Speaker 1: idea was to have a competition. It means at the end,
• 4:06
  Speaker 1: every team is presenting the checker game and they compete
• 4:08
  Speaker 1: against each other and the first one get a price.
• 4:12
  Speaker 1: So where you're coding this and you're basically coding the
• 4:16
  Speaker 1: intelligence behind the checker game or your AI, and you
• 4:20
  Speaker 1: see it improving until the moment it's it's it's better
• 4:23
  Speaker 1: than us, right, And when you see it competing, it's
• 4:27
  Speaker 1: like your kid competing in a tournament.
• 4:30
  Speaker 3: I think it was a very no wall moment because.
• 4:33
  Speaker 1: You create it from scratch and you see it and
• 4:35
  Speaker 1: it's basically winning or losing. It's something that's emotionally very strong, right.
• 4:42
  Speaker 1: And technology technology evolved in the meantime, right because those
• 4:46
  Speaker 1: who are very simple AI are going back at that time,
• 4:49
  Speaker 1: very simple game with limited space. But this whole moment
• 4:53
  Speaker 1: I retrieve it as I basically I grew my career
• 4:56
  Speaker 1: was recently with gen AI and l LMS and stuff
• 4:59
  Speaker 1: like this.
• 4:59
  Speaker 3: But this mom is really a moment I like, and
• 5:02
  Speaker 3: I chose that's a beautiful example. I love it.
• 5:05
  Speaker 2: I love it, and you know, look, welcome again back
• 5:07
  Speaker 2: to the showman. It's nice to have you back.
• 5:09
  Speaker 3: And thanks.
• 5:10
  Speaker 2: Actually, before we before we get into a very technical
• 5:15
  Speaker 2: story we're going to be telling together today, I started
• 5:18
  Speaker 2: to hear more of this year about this term observability
• 5:22
  Speaker 2: as a key term in relation to DEX and back
• 5:25
  Speaker 2: extension Next Thing. Right, what's your understanding of what that
• 5:29
  Speaker 2: term denotes and what key features do you think are
• 5:33
  Speaker 2: a real observability platform in twenty twenty five needs to
• 5:37
  Speaker 2: feature to be able to you know, to to have
• 5:40
  Speaker 2: the right to being called an observability platform.
• 5:44
  Speaker 1: Now, that's a good point, Tom, It's it's also a
• 5:46
  Speaker 1: term I hear more and more also discussing with customers,
• 5:52
  Speaker 1: and it's a term I hear of course, I'll discussing
• 5:54
  Speaker 1: with the customer of nexting, but also next Thing internally.
• 5:57
  Speaker 3: Like engineers and even product EDAs.
• 6:00
  Speaker 1: And to answer your question, I think that there are
• 6:05
  Speaker 1: let's say three important points when it comes.
• 6:07
  Speaker 3: To observability or building an obserbility platform. Right. The first
• 6:11
  Speaker 3: one is simply like.
• 6:14
  Speaker 1: Having the data right, comprehensive data collection that allows you
• 6:18
  Speaker 1: to have the obserability of what's happening and to end right.
• 6:22
  Speaker 1: And if we think at nexting and next in infinity platform,
• 6:25
  Speaker 1: we have it like starting from the device level with
• 6:27
  Speaker 1: the device view, but we also have it we offer
• 6:30
  Speaker 1: it cross customer right with cloud integens, where we basically
• 6:33
  Speaker 1: provide observability across customers with anonymized data. And I think
• 6:38
  Speaker 1: this is a very important point. Let's see, it's a
• 6:40
  Speaker 1: foundation of what we need. And then of course when
• 6:44
  Speaker 1: you have this data, and you have this data at scale,
• 6:47
  Speaker 1: it's hard to make sense out of this data, right
• 6:49
  Speaker 1: because I cannot throw all the data at you and
• 6:51
  Speaker 1: you try to make sense out of it. So we
• 6:53
  Speaker 1: really need to be able to provide let's say insights
• 6:57
  Speaker 1: instead of providing data at scale for our customers. And
• 7:01
  Speaker 1: I think this is coming back to AI with machine
• 7:05
  Speaker 1: learning like simply but also generative AI. There is an
• 7:08
  Speaker 1: opportunity for the end user to be able to make
• 7:11
  Speaker 1: sense out of the data at scale right by just
• 7:14
  Speaker 1: infusing the right insights for the user. So this is
• 7:17
  Speaker 1: the second point, really building insights and not just the data.
• 7:22
  Speaker 1: And finally, I think it's a very important point that
• 7:24
  Speaker 1: closes the loop is actionability right because once you have
• 7:28
  Speaker 1: the data, you have the insights, you have an understanding
• 7:30
  Speaker 1: of what's happening. You need action ability in order to
• 7:33
  Speaker 1: basically remediate or to improve given the situation, right, and
• 7:38
  Speaker 1: this that next thing we have it of course with
• 7:40
  Speaker 1: the remote actions with flows. So I think having the
• 7:44
  Speaker 1: three elements the right data at scale, insights and also
• 7:48
  Speaker 1: the actionability allows us to be able to bring value
• 7:51
  Speaker 1: to our customers and for any obserbility platform to bring
• 7:54
  Speaker 1: value to its users.
• 7:58
  Speaker 4: Mohammad just outlines see diagnosed fix in perfect technical terms, right,
• 8:02
  Speaker 4: we see it, so we can diagnose it with the insights,
• 8:05
  Speaker 4: and we've got the action ability to fix it. So
• 8:07
  Speaker 4: maxim malcom to the show from me as well. For
• 8:10
  Speaker 4: the first time. Before we get into the bit of
• 8:13
  Speaker 4: the development story that we're going to cover, talk a
• 8:16
  Speaker 4: bit about n QL, how it's unique from other query
• 8:21
  Speaker 4: languages in the role that it played here and next
• 8:23
  Speaker 4: thing success.
• 8:27
  Speaker 5: Well, and Curly is a very very important piece in
• 8:32
  Speaker 5: the next few platform. You know, we collect a lot
• 8:34
  Speaker 5: of data for this observability that Muhammed just explained, and
• 8:39
  Speaker 5: we put that in a big data platform and Curly
• 8:42
  Speaker 5: is the unique way for you to talk with that
• 8:46
  Speaker 5: platform and to retrieve the information you need. So you
• 8:50
  Speaker 5: need to know how to write and cur query to
• 8:53
  Speaker 5: get exactly the information you need. But you need also
• 8:55
  Speaker 5: to understand the subtivities of how the data is collected,
• 9:01
  Speaker 5: what data is collected, when the data is invalidated, because
• 9:05
  Speaker 5: we don't keep all the data forever, so you cannot
• 9:08
  Speaker 5: go back in the past as much as you want,
• 9:10
  Speaker 5: for instance, So if you want to be an expert
• 9:15
  Speaker 5: at next Thing and an expert index analysis, you need
• 9:19
  Speaker 5: to know NQL. You need to be an expert at inkl.
• 9:22
  Speaker 5: It's fundamental. It's not only the language, because it's a
• 9:26
  Speaker 5: language that is custom it's not similar to SQL, it's
• 9:31
  Speaker 5: something different. But it's also the data model. You need
• 9:35
  Speaker 5: to understand what QL is used for. So when you
• 9:38
  Speaker 5: request data from a certain table, you need to understand
• 9:42
  Speaker 5: what that table contains, what each column contains. Sometimes we
• 9:47
  Speaker 5: collect metrics, and understanding what each metrics contain is very
• 9:53
  Speaker 5: very important. And CURL is used everywhere, so you can
• 9:58
  Speaker 5: type in directly in Quel to reach data, but you
• 10:01
  Speaker 5: can also have a lot of service retrieve data within
• 10:05
  Speaker 5: the next Thing solution. You retrieve data by writing QL
• 10:10
  Speaker 5: in code, so you have service to service communication, you
• 10:14
  Speaker 5: have prontend to back end communication like the front end,
• 10:18
  Speaker 5: they have a lot of NQL to define what needs
• 10:21
  Speaker 5: to be displayed, for instance. So it's really at the
• 10:24
  Speaker 5: heart of a solution, and if you want to be
• 10:27
  Speaker 5: an expert in your solution, you need to master it.
• 10:31
  Speaker 5: It also has a lot of nice hidden feature because
• 10:35
  Speaker 5: it abstracts away all the complexity in NQL.
• 10:39
  Speaker 6: We can define who can access what. For instance, we
• 10:42
  Speaker 6: have some.
• 10:44
  Speaker 5: Privileged user that can see the names of the employee
• 10:47
  Speaker 5: and these things, but we have other users of NQL
• 10:51
  Speaker 5: that can see a certain user, but its name.
• 10:55
  Speaker 6: Will be hidden, for instance.
• 10:56
  Speaker 5: And all this is abstracted away in permission applied on.
• 11:00
  Speaker 5: So two person typing the same in QL might get
• 11:03
  Speaker 5: slightly different results depending on their permission. And this is
• 11:07
  Speaker 5: a nice feature that Girl brings to abstract away all
• 11:11
  Speaker 5: that complexity.
• 11:14
  Speaker 4: It's fascinating. Through my IT career, I did a little
• 11:18
  Speaker 4: bit of my own SQL querying access databases and SQL
• 11:23
  Speaker 4: server in the back end of SMS to get client data.
• 11:27
  Speaker 4: But it was all self taught. It was all very
• 11:31
  Speaker 4: much trial and error and can be fairly frustrating as well,
• 11:36
  Speaker 4: especially when you're really not confident that you've got the
• 11:38
  Speaker 4: query formed correctly, so you're not confident in the data
• 11:42
  Speaker 4: that you're getting back. Why is it that it teams
• 11:46
  Speaker 4: often struggle with query languages and how is AI beginning
• 11:50
  Speaker 4: to remove some of those roadblocks? So Muhammed, maybe you
• 11:52
  Speaker 4: can jump in here on that one.
• 11:54
  Speaker 3: Yeah, definitely. Tim.
• 11:56
  Speaker 1: First, I want to tell you, I'm impressed. I didn't
• 11:58
  Speaker 1: know you were writing as is in the past, so
• 12:02
  Speaker 1: respect I don't know about you.
• 12:04
  Speaker 2: I'm telling that Muhammad will never heavy end of it.
• 12:08
  Speaker 1: I say that so for me, I know, I mean
• 12:13
  Speaker 1: SQUL was a nightmare for me personally. It's like and
• 12:18
  Speaker 1: I think the first time I had to write SQL
• 12:19
  Speaker 1: during during my studies and we had this class about
• 12:22
  Speaker 1: databases and SQL is powerful and that's why it's the
• 12:26
  Speaker 1: most popular query language out there, of course, but it's
• 12:29
  Speaker 1: very complex, right I remember there's join and outer join
• 12:32
  Speaker 1: and you have.
• 12:32
  Speaker 3: To respect certain things.
• 12:34
  Speaker 1: So it's very it's it's very powerful, but yet very
• 12:39
  Speaker 1: complex and hard to master. Right And when I when
• 12:44
  Speaker 1: I first when I came to Next Thing, when I
• 12:46
  Speaker 1: joined Next Thing three years ago and I discovered the
• 12:49
  Speaker 1: n QL. So now it looks much simpler than SQL.
• 12:52
  Speaker 1: Is great, right, it means I can do it looks
• 12:55
  Speaker 1: very simple and I can do queries. But I'm still
• 12:58
  Speaker 1: you know, I'm see an engineer by training like maxim,
• 13:02
  Speaker 1: and we like to geek out and to discover a
• 13:05
  Speaker 1: new language, even if it's simpler than e SQL. But
• 13:07
  Speaker 1: I think we are not very representative of the overall
• 13:11
  Speaker 1: population because you need to be in the technical part
• 13:14
  Speaker 1: and no scured so and quer remains like a barrier
• 13:18
  Speaker 1: to entry for for certain people to next sink, right,
• 13:23
  Speaker 1: you still need to learn the language.
• 13:25
  Speaker 3: Even if it's simple, you still need to learn it.
• 13:28
  Speaker 3: You seem to need to master to understand the data model.
• 13:30
  Speaker 6: Right.
• 13:31
  Speaker 1: And we hear Pedroals or our CEO also mentioning that
• 13:35
  Speaker 1: we want to democratize DEX. We want to open like
• 13:37
  Speaker 1: next thing to more and more users. And I think
• 13:39
  Speaker 1: if we need to do that, we need to remove
• 13:41
  Speaker 1: any barrier to entry right to next sing. And that's
• 13:47
  Speaker 1: why when we at the inception of assists, right when
• 13:51
  Speaker 1: we started stopped really as like as as a way
• 13:55
  Speaker 1: to learn NQUL, we want to say, what about there
• 13:58
  Speaker 1: is an opportunity today with generative AI, those models where
• 14:00
  Speaker 1: you can explain in a simple sentence what you want
• 14:04
  Speaker 1: and then get it. Let's say, translated in another language.
• 14:08
  Speaker 1: Can be NQL, but can be anything else. So there
• 14:10
  Speaker 1: is an opportunity for us to simplify the user experience, right,
• 14:14
  Speaker 1: And I think the best, let's say, best query language
• 14:20
  Speaker 1: is English. So if our users are able to write
• 14:22
  Speaker 1: their question in English and get the right data they need,
• 14:26
  Speaker 1: I think we'll have done a great job in really
• 14:28
  Speaker 1: democratizing decks and access to our to nexting.
• 14:32
  Speaker 4: Right, So we'll explain on that a little bit. I
• 14:36
  Speaker 4: can't imagine, so I don't have that grasp of the
• 14:40
  Speaker 4: knowledge required to train a language model. But language is
• 14:46
  Speaker 4: in general, whether it's spoken language or a query language,
• 14:49
  Speaker 4: they're very specific to syntax and structure. The query language
• 14:53
  Speaker 4: is also you've got to get the commands correct. And
• 14:56
  Speaker 4: like you said that all of the structural elements of JOINT,
• 15:00
  Speaker 4: did that help or hinder the teams when training assists
• 15:05
  Speaker 4: to generate NQL from just a simple natural language question?
• 15:12
  Speaker 6: Yeah.
• 15:12
  Speaker 5: Indeed, when we started, I mean we wanted to use
• 15:17
  Speaker 5: commercial lms to do that, and that was one of
• 15:20
  Speaker 5: the quick finding is that they don't know NQL, so
• 15:24
  Speaker 5: they make a lot of syntax mistakes because they think
• 15:28
  Speaker 5: they can translate what they know from SQL to NQL
• 15:32
  Speaker 5: easily and they start to assume that the.
• 15:36
  Speaker 6: Language will work similarly. But that's not the case.
• 15:40
  Speaker 5: You know, these LM they are trained by just ingesting
• 15:44
  Speaker 5: all the information they can get from the Internet, and
• 15:48
  Speaker 5: this is one of the reasons they are good at
• 15:51
  Speaker 5: very known code language like Python, Java, C plus plus
• 15:58
  Speaker 5: Yeah or squel, because there's a lot of it on
• 16:02
  Speaker 5: that It's discussed on stack overflow extensively, it's shared on
• 16:08
  Speaker 5: GitHub publicly, so you have a lot of sources of
• 16:11
  Speaker 5: information for all the language. If you compare that to NQL,
• 16:16
  Speaker 5: it's very very tiny. It's only present in the documentation
• 16:19
  Speaker 5: of the next Thing platform and even though it has
• 16:22
  Speaker 5: some similarities with other language. Originally it was based on
• 16:26
  Speaker 5: a language by Microsoft named Cusco.
• 16:31
  Speaker 6: It evolved and it's no longer like Cusco.
• 16:33
  Speaker 5: You cannot use Cusco and say I know couscos right,
• 16:37
  Speaker 5: QUL that doesn't work. We have additional statement, we have
• 16:40
  Speaker 5: additional constraint, and then you have the data model. So
• 16:47
  Speaker 5: the foundational knowledge of the Vanila LM doesn't contain NQL.
• 16:54
  Speaker 6: It was not optimized for that.
• 16:55
  Speaker 5: Because the optimized to work for most of the use case,
• 16:59
  Speaker 5: and NQL is a very very limited your case in fact,
• 17:03
  Speaker 5: so if you say to an M two to CHANGEPT
• 17:07
  Speaker 5: for instance, to generate QL, it might seem to work.
• 17:11
  Speaker 5: After a few attempts, you might have it. Okay, it
• 17:13
  Speaker 5: seemed to work, But when you dig a bit deeper,
• 17:16
  Speaker 5: it will hallucinates very very often. So it will invent colony,
• 17:21
  Speaker 5: you will invent statement, you will invent structures that are
• 17:24
  Speaker 5: not correct. So when we started, this was the big challenge.
• 17:30
  Speaker 5: And another one is okay, if you want to if
• 17:35
  Speaker 5: you if you want to fill that gap of knowledge,
• 17:37
  Speaker 5: then you need to provide it with data. But we
• 17:40
  Speaker 5: didn't have data when we started. We were like, okay,
• 17:43
  Speaker 5: where can we get data to show enough NQL queries
• 17:48
  Speaker 5: so that it starts to grasp how the language is structured?
• 17:52
  Speaker 5: What are the constraints. So these are the two main
• 17:57
  Speaker 5: challenges when we started.
• 17:59
  Speaker 1: Yes, it's like if you think of it's like like humans,
• 18:04
  Speaker 1: right as those like those pre training the l l ms,
• 18:07
  Speaker 1: they basically they have seen so much on the web.
• 18:10
  Speaker 1: They have to say the average knowledge, but they don't
• 18:13
  Speaker 1: know when q el and it's like we need to
• 18:15
  Speaker 1: somehow train them, right, we need training them is not
• 18:19
  Speaker 1: showing them one or two examples. It's like really training them.
• 18:23
  Speaker 1: Like our users are getting training in the NQLS. We
• 18:25
  Speaker 1: need to train the l l M and assist in
• 18:27
  Speaker 1: order to produce good quality and relatively complex and co queries.
• 18:34
  Speaker 2: Is what we call going from checkers to chess, Muhammad,
• 18:37
  Speaker 2: Yes exactly, that's.
• 18:41
  Speaker 4: Maybe touch on that journey right. It's because from a
• 18:45
  Speaker 4: new term for me retrieval, augmented generation to fine tuning.
• 18:50
  Speaker 4: We're really taking an iterative approach to this versus a
• 18:53
  Speaker 4: big bang, right, talk us through the journey of how
• 18:57
  Speaker 4: we're we're starting with something really good good and then
• 19:01
  Speaker 4: evolving it to something great.
• 19:03
  Speaker 3: Yeah, I think when.
• 19:06
  Speaker 1: We when we when we we we started assist as
• 19:12
  Speaker 1: you said, it's really I mean a journey, and we
• 19:14
  Speaker 1: started like with with.
• 19:16
  Speaker 3: With like.
• 19:18
  Speaker 1: A simple model, and then we went incrementally in order
• 19:21
  Speaker 1: to to to improve as we basically as we as
• 19:25
  Speaker 1: we as we go and how did we do it?
• 19:28
  Speaker 3: First thing?
• 19:28
  Speaker 1: And as usually this is what we do when it
• 19:30
  Speaker 1: develop an AI model in general, either assist on any
• 19:35
  Speaker 1: other AI model is really to say okay, to start
• 19:38
  Speaker 1: with a benchmark, right, because the benchmark basically will allow
• 19:41
  Speaker 1: us will be like the north star and will allow
• 19:43
  Speaker 1: us to to see whether we're going in the right
• 19:46
  Speaker 1: direction or not. Right, so we we really started with
• 19:50
  Speaker 1: a benchmark that is basically, in simple terms, user questions
• 19:57
  Speaker 1: and right.
• 19:57
  Speaker 3: Let's say, que queries associated with these user questions.
• 20:02
  Speaker 1: And once we call it actually the golden data set,
• 20:04
  Speaker 1: that is basically our north star, allowing us to see
• 20:07
  Speaker 1: how good we are performing. And once we had this
• 20:10
  Speaker 1: in place, then we can start like we started very
• 20:14
  Speaker 1: pragmatically with like first what you called zero short learning.
• 20:19
  Speaker 1: It means asking, as maxim said, LM to generate an
• 20:23
  Speaker 1: n COL query, but just explaining what is n col
• 20:26
  Speaker 1: and providing some information about data model. Very quickly we
• 20:29
  Speaker 1: saw that we needed by injecting relevant examples in the
• 20:34
  Speaker 1: context what we call the context of the LLM, think
• 20:35
  Speaker 1: of it as a memory.
• 20:36
  Speaker 3: Or the ram of the LLM.
• 20:39
  Speaker 1: By providing the right examples that are associated with the
• 20:41
  Speaker 1: question the user is asking, the performance improved. And this
• 20:46
  Speaker 1: is what we call like retrieval ofmental generation or RAG
• 20:50
  Speaker 1: retrieval in the sense if you ask a question Tom,
• 20:54
  Speaker 1: the algorithm retrieve questions that are similar, right, and inject
• 20:59
  Speaker 1: the answer in the memory of the LLM.
• 21:02
  Speaker 3: Right.
• 21:03
  Speaker 1: So it's a quite elegant way in order to provide
• 21:07
  Speaker 1: in additional to the memory what the LM learned in
• 21:09
  Speaker 1: the past, giving something fresh that is relevant to the
• 21:13
  Speaker 1: user question.
• 21:13
  Speaker 6: It's kind of a.
• 21:14
  Speaker 5: Quick patch, right, Okay, yeah, I know you don't know that,
• 21:17
  Speaker 5: but that we'll quickly squeeze it before you answer so
• 21:21
  Speaker 5: you can read.
• 21:21
  Speaker 6: The right answer and hopefully you can do a better job.
• 21:24
  Speaker 1: Exactly, And after we saw the drug that we improve
• 21:28
  Speaker 1: the performance and perhaps maximum is like the expert can
• 21:32
  Speaker 1: elaborate on the stages afterwards, where we really I mean
• 21:37
  Speaker 1: took it to the next level with fine tuning.
• 21:40
  Speaker 5: Right, Yeah, I guess at some point we realized that
• 21:43
  Speaker 5: we started to have a lot of data in that
• 21:47
  Speaker 5: database of RAG that at some point we could maybe
• 21:52
  Speaker 5: directly inject not only the two or three closest examples
• 21:58
  Speaker 5: and show them to the LLLM, but inject all of
• 22:01
  Speaker 5: those and it's hundreds of thousands directly in the memory.
• 22:06
  Speaker 5: And for that we need to process to what we
• 22:08
  Speaker 5: call the fine tuning, which is basically taking a model
• 22:13
  Speaker 5: that exists and in our case we took cord Lama,
• 22:16
  Speaker 5: which is a model specialized for code, and we start
• 22:19
  Speaker 5: to show it examples, like we show it a batch
• 22:23
  Speaker 5: of examples where this is the question what do you produce?
• 22:27
  Speaker 5: And at first it will produce queries that are probably
• 22:31
  Speaker 5: like seql and so on.
• 22:34
  Speaker 6: But by.
• 22:37
  Speaker 5: Training, the model is just updating the prediction so that
• 22:40
  Speaker 5: the next time it does it a little bit better.
• 22:43
  Speaker 5: And so we do it by badge and anytime we
• 22:46
  Speaker 5: update the weights, so the weights is the memory of
• 22:49
  Speaker 5: the of the network.
• 22:51
  Speaker 6: Then it gets better and better at producing.
• 22:53
  Speaker 5: At some point we see less and less syntax her
• 22:57
  Speaker 5: At some point it starts so to learn, okay, how
• 23:00
  Speaker 5: to relate what the user.
• 23:02
  Speaker 6: Asked to what needs to be in the output.
• 23:06
  Speaker 5: And this fine tuning process is something that is a
• 23:10
  Speaker 5: bit resource intence. I you need gpused to do that.
• 23:13
  Speaker 5: As I said, you need a lot of data. But
• 23:15
  Speaker 5: once that is done, then you have your own model
• 23:18
  Speaker 5: and you can you can prompt it with very much
• 23:22
  Speaker 5: simply right. You just give it the question and then
• 23:24
  Speaker 5: it knows NQL and it's much better.
• 23:29
  Speaker 6: Because you can you can give it the whole.
• 23:35
  Speaker 5: Varieties in all the tables. It has much more examples
• 23:41
  Speaker 5: that it can use from its memory to predict what
• 23:44
  Speaker 5: would be the best query to answer. So this is
• 23:49
  Speaker 5: what we could find tuning. We don't need to have
• 23:52
  Speaker 5: a specialized model for NQL that is interacting with the
• 23:57
  Speaker 5: other more come on ms that taking chart to chat
• 24:03
  Speaker 5: part of assist, for instance.
• 24:06
  Speaker 4: It's really fascinating to hear the journey of learning and
• 24:10
  Speaker 4: quel and then fine tuning it based on some of
• 24:14
  Speaker 4: the answers it gets from the queries that are generated,
• 24:18
  Speaker 4: and then being able to take action on the things
• 24:21
  Speaker 4: that it learned. It's really that's kind of what I
• 24:22
  Speaker 4: want to put to bed Muhammad here is if I
• 24:25
  Speaker 4: were to walk away now, the perception I have is
• 24:28
  Speaker 4: this is really just a query generator, and it's so
• 24:32
  Speaker 4: much more than that. It's really evolving into a true
• 24:36
  Speaker 4: dex analyst, a collaborative partner. So beyond just simplifying NQL,
• 24:43
  Speaker 4: how does assist really enhance it operations from troubleshooting specific
• 24:49
  Speaker 4: issues all the way through strategic decision making. Let's make
• 24:53
  Speaker 4: sure folks aren't left with the thought that this is
• 24:56
  Speaker 4: just generating queries.
• 24:58
  Speaker 1: It's a It's a very good point, and our vision
• 25:03
  Speaker 1: for ASSISTS is really to be.
• 25:07
  Speaker 3: To be.
• 25:10
  Speaker 1: A dex analyst for any user that is using our
• 25:13
  Speaker 1: platform and Nexing Infinity. And this dex analyst is is
• 25:18
  Speaker 1: getting better and better, getting more seniority, more experience, and
• 25:22
  Speaker 1: also having access to more tools. Right, because think of it.
• 25:26
  Speaker 1: In the beginning, we start assist mainly with a tool
• 25:30
  Speaker 1: that is generating as you say, like quel queries, but
• 25:32
  Speaker 1: this is not the vision.
• 25:33
  Speaker 3: It just to start right.
• 25:34
  Speaker 1: The idea is to provided more tools, more powers in
• 25:37
  Speaker 1: order to be able to help the user. Right, And
• 25:41
  Speaker 1: our vision is really to have a true Decks analyst
• 25:44
  Speaker 1: that is helping you in the product getting the right
• 25:47
  Speaker 1: insights you need. And the tools are like generating queris
• 25:51
  Speaker 1: as you said, but now it's basically generating engage campaigns
• 25:54
  Speaker 1: based on a simple description, getting answer from the documentation,
• 25:59
  Speaker 1: and now it's really about also making analysis right, so
• 26:02
  Speaker 1: basically compiling different sources, compiling investigation plus some search and
• 26:08
  Speaker 1: the product documentation and basically providing you with an actual
• 26:13
  Speaker 1: analysis that is related to the question.
• 26:15
  Speaker 3: Right.
• 26:18
  Speaker 1: And Maxim very recently did a demo with some very
• 26:21
  Speaker 1: impressing about like decks to that OD that we released
• 26:24
  Speaker 1: recently and you can see now you can also ask
• 26:28
  Speaker 1: open ended question questions that are not very well defined,
• 26:31
  Speaker 1: and assists basically will go and look in different tools
• 26:35
  Speaker 1: and the pages of Next Thing and certain dashboard and
• 26:38
  Speaker 1: documentation and investigation and provide you with insights that are
• 26:41
  Speaker 1: related to the question.
• 26:42
  Speaker 3: So it's very impressing. This is a vision we're having
• 26:45
  Speaker 3: for assists for our.
• 26:46
  Speaker 5: Users, and that's something even just with NQL. When you
• 26:50
  Speaker 5: build a text two QL system, you have one question,
• 26:53
  Speaker 5: one query, but often we realize people they ask a
• 26:57
  Speaker 5: questions that would require multiple queries to answer properly, and
• 27:01
  Speaker 5: as is true, although it is capable of doing that,
• 27:03
  Speaker 5: it is capable of saying okay for that question, I
• 27:06
  Speaker 5: would need to run several investigations to give a proper ranswer,
• 27:12
  Speaker 5: not even mentioning the other tools that we are putting
• 27:14
  Speaker 5: in the documentation of the other things, but just saying okay,
• 27:18
  Speaker 5: I need at least two or three queries to give
• 27:21
  Speaker 5: a complete response to that user.
• 27:25
  Speaker 6: This is something that we are going towards.
• 27:28
  Speaker 1: And to add one last point, it's also aligned with
• 27:30
  Speaker 1: the vision we have at Next Thing in the sense
• 27:33
  Speaker 1: building agents that are very specialized and one that are
• 27:39
  Speaker 1: experts and specialized in one domain that are able to
• 27:42
  Speaker 1: help the user. Assist is basically the decks analysts the expert,
• 27:46
  Speaker 1: and next Inc Allowing the user to get the inside
• 27:49
  Speaker 1: they need. Autopilot is basically the expert in it and
• 27:52
  Speaker 1: it resolution allowing our users to fix it issues. It's
• 27:56
  Speaker 1: really the vision we've having with a very strong specialized
• 28:00
  Speaker 1: is that we train and that basically are ramping up
• 28:03
  Speaker 1: and speed and improving right starting as I know, an
• 28:05
  Speaker 1: intern and then engineer and then an expert, and then
• 28:08
  Speaker 1: that's the vision we have.
• 28:10
  Speaker 4: It's fascinating. Before I give it back to Tom, just
• 28:12
  Speaker 4: a comment. Just before recording this podcast, I was on
• 28:16
  Speaker 4: I did a live session with probably fifteen new hires,
• 28:20
  Speaker 4: new brand new Next sneakers, and my session was all
• 28:23
  Speaker 4: about the evolution of end user compute from the level
• 28:26
  Speaker 4: one helpdesk technician to the level two desktop support analysts,
• 28:30
  Speaker 4: the level three desktop engineering and level four or the
• 28:34
  Speaker 4: enterprise architect. The different skills and knowledge and expertise that
• 28:38
  Speaker 4: they gather over time, the different tools they have access to,
• 28:41
  Speaker 4: the different privileges they have, and it's fascinating to see
• 28:44
  Speaker 4: the work that you're doing parallel that journey in a
• 28:48
  Speaker 4: digital way to what a human would do, and it's
• 28:51
  Speaker 4: over the course of its career, far faster and with
• 28:54
  Speaker 4: some really great results. It's just really interesting timing for
• 28:57
  Speaker 4: me personally, having just done that session and then hear
• 29:00
  Speaker 4: you talk about that same progression here. It is really cool, really.
• 29:03
  Speaker 3: Cool, parallel, amazing story.
• 29:07
  Speaker 2: Guys, any surprising or creative ways that it's already been
• 29:10
  Speaker 2: leveraged out there by by customers and users to solve
• 29:13
  Speaker 2: operational challenges that you could share.
• 29:18
  Speaker 5: I think we had some surprise at the beginning, maybe
• 29:24
  Speaker 5: not that they were particularly creative, but we were surprising
• 29:29
  Speaker 5: the way they were using it because, and this is
• 29:31
  Speaker 5: a lesson learned, they typically don't explain well what they need.
• 29:36
  Speaker 5: They explained a broad problem that they have, and the
• 29:40
  Speaker 5: first version of assist would take one querity once for
• 29:43
  Speaker 5: that they would while we were expecting them to explain
• 29:48
  Speaker 5: clearly what they need, they would say, oh, this species
• 29:50
  Speaker 5: is slow.
• 29:51
  Speaker 6: Why uh?
• 29:52
  Speaker 5: And for from from a pure enterance perspective, like a
• 29:57
  Speaker 5: slow pecs many things it's is is it slow to boot?
• 30:02
  Speaker 5: Is it slow to load a page on the internet?
• 30:05
  Speaker 5: Is it slow to make a web request? All these things?
• 30:09
  Speaker 5: So we had to learn a bit and step back
• 30:12
  Speaker 5: and say, Okay, they or users will keep asking these question,
• 30:16
  Speaker 5: so how can we address them better. Another surprising thing
• 30:20
  Speaker 5: that we saw that they started to use a syst
• 30:24
  Speaker 5: to ask quite basic question, not necessarily because they don't
• 30:28
  Speaker 5: know the answer, but because it's quite fast to get
• 30:31
  Speaker 5: it through assists. So they went there the past a
• 30:35
  Speaker 5: list of devices, for instance, and they say, okay, give
• 30:38
  Speaker 5: me the last good time of all these devices, so
• 30:41
  Speaker 5: they can quickly extract information. And I don't know exactly
• 30:45
  Speaker 5: what was the process, but they have this way of
• 30:50
  Speaker 5: working that we try to adjust to and serve them
• 30:53
  Speaker 5: better all.
• 30:54
  Speaker 6: These use cases.
• 30:58
  Speaker 2: And with that kind of evolution into into a true
• 31:02
  Speaker 2: AI AI agent, quite difficult thing to say. What future
• 31:08
  Speaker 2: capabilities are you most excited about? And what future milestone
• 31:11
  Speaker 2: should listeners be aware about about what's coming next.
• 31:14
  Speaker 3: With us this?
• 31:18
  Speaker 1: Yeah, I think for what's coming next to assist let
• 31:21
  Speaker 1: maximum answer afterwards. But there are two things I'm excited
• 31:25
  Speaker 1: about personally, and there are opportunities and our for next thing.
• 31:32
  Speaker 1: First thing is what we call reasoning models. Now I
• 31:35
  Speaker 1: don't know if you're using gpt O one. We also
• 31:38
  Speaker 1: recently heard about deep sick one. Those are the models
• 31:41
  Speaker 1: that are we call reasoning models. Basically, they take more
• 31:46
  Speaker 1: time to to to Basically they verbalize their votes, right,
• 31:51
  Speaker 1: and they take more time before answering. And this we've
• 31:55
  Speaker 1: seen in the different baschmark and the research community that
• 31:57
  Speaker 1: is improving the quality of the output. Right, It's like
• 31:59
  Speaker 1: you and you know if I ask you a question,
• 32:01
  Speaker 1: I give you one second to answer the question, will
• 32:06
  Speaker 1: the answer.
• 32:06
  Speaker 3: Will not be as great as giving you one minute.
• 32:08
  Speaker 1: Right, you go for one minute and you do some
• 32:10
  Speaker 1: research and you come back to me, Tom, Right, I
• 32:12
  Speaker 1: know you're.
• 32:12
  Speaker 2: Very exactly it's the same Mohammed.
• 32:14
  Speaker 5: But yeah, and I think when you read the thinking
• 32:17
  Speaker 5: process of one's really impressive because you seem to read
• 32:21
  Speaker 5: the mind of someone that thinks very fast, makes assumption,
• 32:26
  Speaker 5: corrects itself and come back to the previous assumptions. And
• 32:29
  Speaker 5: no way that should do this because he asked that,
• 32:32
  Speaker 5: but given this, maybe I should do this way.
• 32:35
  Speaker 6: Let's step back. It's really it's really amazing to read.
• 32:40
  Speaker 1: And what's fascinating for me and just is like the
• 32:44
  Speaker 1: way they train it, right, is like reinforcement learning. So
• 32:47
  Speaker 1: they didn't when they trained this, Lamb, they didn't ask
• 32:50
  Speaker 1: to think. It's basically to learn through like reinforcement learning
• 32:54
  Speaker 1: by seeing examples and failing, but basically thinking longer will
• 32:59
  Speaker 1: allow us to have more accurate answers. Right, That's what's
• 33:01
  Speaker 1: fascinating and for us, I think there is an opportunity
• 33:05
  Speaker 1: for everything we're doing right for like nexting assist, but
• 33:08
  Speaker 1: also autopilot to fix it issues. I think we can
• 33:11
  Speaker 1: capitalize all those reasoning models that are taking more time
• 33:14
  Speaker 1: to produce the output in order to have a higher
• 33:17
  Speaker 1: quality outcome and value for our customers. If we think
• 33:21
  Speaker 1: as we discussed with dex analysts, for example, we can
• 33:24
  Speaker 1: imagine having asking a very complex question to assist very advanced,
• 33:30
  Speaker 1: and then assists can go on reasoning mode, tell you wait,
• 33:33
  Speaker 1: give me two minutes and I will get you a
• 33:35
  Speaker 1: report about what you're asking, and then going there reasoning,
• 33:38
  Speaker 1: fetching different sources and giving you something very strong right
• 33:42
  Speaker 1: wrong around this very strategic question that would require I know,
• 33:46
  Speaker 1: two or three days for human right.
• 33:48
  Speaker 3: So this aspect is very interesting for me. It's very fascinating.
• 33:52
  Speaker 1: Right, And the intellectual site and the second thing I'm
• 33:58
  Speaker 1: curious about I would like to also push is like,
• 34:01
  Speaker 1: really this multi agent collaboration right there is a lot
• 34:04
  Speaker 1: of research about multi agent collaboration, where basically you have
• 34:07
  Speaker 1: different agents that are specialized and that co operate in
• 34:11
  Speaker 1: order to achieve a common goal. Right now, we would
• 34:15
  Speaker 1: love Maximum and myself to discover to explore if we
• 34:18
  Speaker 1: take next Assist as the expert and next thing and
• 34:21
  Speaker 1: also Autopied as the expert in it and IT issue resolution,
• 34:25
  Speaker 1: if we make them work together, I think we can
• 34:27
  Speaker 1: achieve something very, very strong and beautiful for our customers.
• 34:30
  Speaker 3: So this is like something I would love to explore further.
• 34:34
  Speaker 6: Definitely, I agree.
• 34:36
  Speaker 3: I think maxim you can end with us masters things.
• 34:39
  Speaker 2: Things the listeners should be aware of is what's coming
• 34:42
  Speaker 2: next for assist concretely.
• 34:45
  Speaker 5: Well, I think we have we have short term goals
• 34:47
  Speaker 5: and we have long term goals. Short term is like
• 34:51
  Speaker 5: we are we are adding more and more of the
• 34:54
  Speaker 5: sources that Moham mentioned, but we are also making the
• 34:58
  Speaker 5: output of assists more actionable and concretely, one thing we
• 35:03
  Speaker 5: would like to do is to be able to generate
• 35:06
  Speaker 5: plots and something that we will soon enable for some
• 35:12
  Speaker 5: better customers, they can basically explain what they want and
• 35:16
  Speaker 5: instead of having a description of the number, the assist
• 35:21
  Speaker 5: could directly make a visualization of it, and the result
• 35:25
  Speaker 5: is fascinating because you ask a broad question and you
• 35:28
  Speaker 5: get a mini report with two three plots and an
• 35:32
  Speaker 5: explanation of the interpretation of each. It is very powerful,
• 35:36
  Speaker 5: I think, and this will be enable predition. Another thing
• 35:43
  Speaker 5: we are working on is to be able to add
• 35:45
  Speaker 5: more knowledge base. Typically when people ask, okay, what is
• 35:49
  Speaker 5: the reason for that problem or what is the root
• 35:52
  Speaker 5: cause of that problem? I observe in the data as
• 35:56
  Speaker 5: it does not necessarily know that well, but we could
• 36:00
  Speaker 5: populate some knowledge base of Okay, this problem typically is
• 36:05
  Speaker 5: a driver problem, or this version of this software is
• 36:11
  Speaker 5: known to produce these kind of bugs, some very punctual
• 36:15
  Speaker 5: knowledge that we could inject it injecting in the memory
• 36:20
  Speaker 5: in the short term memory of Assist, but also customer
• 36:25
  Speaker 5: knowledge base, like they could say, I find I find
• 36:28
  Speaker 5: this resource useful, and I wanted ceast to be aware,
• 36:32
  Speaker 5: so someone in a company could put it into assist
• 36:35
  Speaker 5: and all the colleagues, all these colleagues.
• 36:37
  Speaker 6: Could benefit from it.
• 36:39
  Speaker 5: Whenever he asked, okay, I have this problem with a device,
• 36:42
  Speaker 5: ASSIST could say, oh, your colleague also had this problem
• 36:45
  Speaker 5: and made this recommendation of what is the.
• 36:48
  Speaker 6: Resolution plant.
• 36:51
  Speaker 5: And then more long term what Mohammed described like the
• 36:57
  Speaker 5: true analyst mode where you could could draw a problem
• 37:00
  Speaker 5: at as system and then it would go away for
• 37:03
  Speaker 5: a few minutes but come back with a consolidated report
• 37:07
  Speaker 5: that contains the results of multiple investigation.
• 37:11
  Speaker 1: Uh.
• 37:11
  Speaker 5: And because it had to crunched the data in certain
• 37:15
  Speaker 5: direction analyzing the issue you describe, and so on.
• 37:19
  Speaker 2: Maybe and actually love for the line Muhammad, from from
• 37:24
  Speaker 2: teaching AI checkers to teaching all of these crazy, crazy,
• 37:29
  Speaker 2: much more complex languages and continuing to hear that kind
• 37:32
  Speaker 2: of joy as you watch AI step up to every
• 37:36
  Speaker 2: learning task that you give it and continue to kind
• 37:39
  Speaker 2: of amaze you through your through your career. Real pleasure
• 37:44
  Speaker 2: to have you both on and and hear all about that.
• 37:47
  Speaker 2: Thank you so so much, guys.
• 37:49
  Speaker 3: Thank you, Thank you for having us a preasure.
• 37:51
  Speaker 4: Nothing else I learned today Who to call when I'm
• 37:54
  Speaker 4: struggling with my next sequel query.
• 37:58
  Speaker 2: Simply will have links in the show notes to to
• 38:02
  Speaker 2: that blog that you co authored as well for a
• 38:04
  Speaker 2: little bit more info and uh and of course to
• 38:07
  Speaker 2: h to some more information about assistful interested listeners, which
• 38:12
  Speaker 2: I hope there are a few about that. Okay, thank
• 38:15
  Speaker 2: you so so much, gentlemen, have a great day.
• 38:17
  Speaker 6: Thank you, thank you.
• 38:18
  Speaker 3: Bye bye bye bye