HH56-28-10-21_mixdown.mp3-from OneDrive

Harpreet: [00:00:09] What's up, everybody, welcome, welcome to the artist of Data Science, Happy hour, No. 56, I believe this is 56 happier hours in a row. Man, we've been out here just dishing out wonderful information and helping people out. Shout out to everybody in the room. Christian is back. Good to see you again, Christian. Good to see you. Indeed. Eric, what's going on with the 3D glasses managing the building? Gina, good to see you. Monica Royal is in the building to hopefully AIs had a good good week. Hopefully got a chance to tune into the episode I released

Harpreet: [00:00:40] Today with the

Harpreet: [00:00:42] One and only legendary Andy Hunt.

Harpreet: [00:00:45] He is coauthor

Harpreet: [00:00:47] Of the pragmatic programmer and many, many other books as well. I've been going live a lot this week, I went live, I went live with Danny Ma yesterday, so that was a lot of fun. I went live with

Harpreet: [00:01:00] Another legendary

Harpreet: [00:01:01] Person, an Oxford professor, Marcus Du So toy. He's written the creativity code, music of the primes, he's written this the newest book is called Thinking Better. So I had a great opportunity to chat with with him, we talked a lot about just kind of his philosophy of mathematics and what he loves about mathematics and creativity and things like that. It was a really great conversation. I really enjoyed it. He was also kind enough to introduce me to Dr David Spiegelhalter, so hopefully I can get Dr David Spiegelhalter on the show. He wrote the book The Art of Statistics How to Learn from Data. That's right. Cambridge Look at this. Look at me here from the St Sacramento, California in the hood, and I'm talking with Oxford professors like this mind boggling. And that is that it's pretty cool. What else do I do? I spoke to daily on Hulu this week. We did not go live, though. We just had a little bit and that will be released at some point in the future. So [00:02:00] looking forward to releasing that for you all. And Yemen, hopefully you guys are excited for the weekend, man, this entire month I went, I counted. I was live over 20 times over 20 times, including, you know, presenting a dedicated mic at oDesk twice and the houseless that comet gone live, man, it's been one hell of a month.

Harpreet: [00:02:23] I hope you

Harpreet: [00:02:23] Guys are excited for Halloween.

Harpreet: [00:02:26] I, for one,

Harpreet: [00:02:27] Am indifferent to

Harpreet: [00:02:28] That. But I like seeing

Harpreet: [00:02:30] Little cute kids in their costumes, so there'll be a lot of fun and you know, we can

Harpreet: [00:02:34] Dress my son up like a like a cow. That's crazy, man. I want to go try

Harpreet: [00:02:40] To put him to bed earlier today, and he was just

Harpreet: [00:02:42] Screaming, screaming.

Harpreet: [00:02:44] I had to drive over to the in-laws house

Harpreet: [00:02:45] To kind of

Harpreet: [00:02:46] Put him sleep, and he screamed so much that he vomited

Harpreet: [00:02:50] And he vomited all over me. The kids in St

Harpreet: [00:02:55] Say kids, they have. It'll be fun. It's not, I know it is. It's awesome. I love that kid. But yes, hopefully

Harpreet: [00:03:01] You guys are excited to be

Harpreet: [00:03:02] Here. If you have questions, go ahead. Let me know. Drop your questions in the chat. I see you are all enjoying on LinkedIn and on YouTube.

Harpreet: [00:03:10] So I want to kick

Harpreet: [00:03:11] This thing off.

Harpreet: [00:03:12] I made a post just a little bit

Harpreet: [00:03:13] Earlier today, a couple of hours ago.

Harpreet: [00:03:15] It's not that I might trigger some people

Harpreet: [00:03:17] And kick off a discussion, and that post was Data. Engineering is the most

Harpreet: [00:03:23] Important part

Harpreet: [00:03:24] Of Data science. I want to get your reactions to that statement. I use that. I use the meme taking a page out of Daniel's book. I used that, you know, was it changed my mind? Mean, so yeah. So is Data Engineering the most important part of Data science? I'd love to hear what you guys think. Let's start with with sort of Christian has been a while since we since we heard from you and then we can

Harpreet: [00:03:50] Go to Eric

Harpreet: [00:03:51] And then Matt Blaser is in the building. And if you guys got questions on anything whatsoever, please do let me know if you guys got questions, I will gladly [00:04:00] take them.

Speaker3: [00:04:01] Christian, go for it. Yeah, no, I saw that. I saw that post right before I jumped in and ask you it would be a good discussion right there. I'm actually interested to hear from folks who are sitting in the seat right now doing it and interested to hear from Eric and some of these other guys on whether or not that's true. I mean,

Harpreet: [00:04:14] From our standpoint, there's one

Speaker3: [00:04:16] Thing that I've heard several times and it's like, Hey, if you are doing stuff standalone and Python notebooks and with clean curated data sets, you really haven't gotten down to the arena yet, right? So being able to have that data at your fingertips seems pretty crucial. So I'm looking forward to hearing answers. Eric, you're in the battlefield.

Harpreet: [00:04:33] What do you what do you think? Eric, fear for me. Yeah, so my first

Speaker4: [00:04:39] Thought was,

Harpreet: [00:04:41] Oh heck yeah, like, I

Speaker4: [00:04:43] Am grateful every day that I'm not a

Harpreet: [00:04:44] Data engineer, but I'm also

Speaker4: [00:04:46] Grateful every day that someone is because it's work that

Harpreet: [00:04:51] I personally, I don't get jazzed

Speaker4: [00:04:53] About it. Sometimes I get jazzed about little pieces of it.

Harpreet: [00:04:57] But then I

Speaker4: [00:04:57] Think I'm going to read about Data engineering, and I get like one paragraph into a blog post and I'm falling asleep. So it's just not. I just don't get excited about it. But I know it's super important because any time I want to make a cool change or something in my work. It comes down to some pipeline somewhere needs

Harpreet: [00:05:17] To be adjusted,

Speaker4: [00:05:19] And so I know it's super important. On the other hand, though, as I tried to

Harpreet: [00:05:24] Think about whether or not

Speaker4: [00:05:26] It's really the foundation

Harpreet: [00:05:27] You, I think

Speaker4: [00:05:28] That you can kind of

Harpreet: [00:05:30] Quote unquote, you can

Speaker4: [00:05:31] Outsource some

Harpreet: [00:05:32] Data engineering on a small

Speaker4: [00:05:34] Scale

Harpreet: [00:05:35] Because a few

Speaker4: [00:05:37] A few months ago, my partner and I were working on a small project for like a

Harpreet: [00:05:43] Nonprofit.

Speaker4: [00:05:44] And they're small, they're not big, they don't have a Data engineer. They don't they

Harpreet: [00:05:49] Only have a two

Speaker4: [00:05:50] Employees, right? But they have various services and things like that that they are subscribed to.

Harpreet: [00:05:56] And we were able to bring

Speaker4: [00:05:58] Data

Harpreet: [00:05:58] Together manually from [00:06:00] different

Speaker4: [00:06:00] Sources to get real business results like make

Harpreet: [00:06:04] Change. But we didn't necessarily need a Data engineer

Speaker4: [00:06:08] In order to facilitate any of that happening.

Harpreet: [00:06:10] So I just think that it's just like,

Speaker4: [00:06:11] I think it's just with scale. It does really become foundational. But if you're doing something on a small scale, never

Harpreet: [00:06:18] Underestimate the power of, you know, one person with willpower and

Speaker4: [00:06:23] Enough enough time to figure it out.

Harpreet: [00:06:25] Eric, thank you so much. Let's go to let's go to Monica, and then after Monica, we will go to Matt Plaza.

Speaker5: [00:06:32] First of all, I just want to say that I think I'm way too overly excited for Halloween.

Harpreet: [00:06:38] I love it. I absolutely love it. Is that yeah, you got to break this down for us. I don't immediately get that as Albert Einstein.

Speaker5: [00:06:47] So I'm a data scientist,

Harpreet: [00:06:51] A mad scientist. I love it.

Speaker5: [00:06:53] I love it. But I also

Harpreet: [00:06:54] Like the most

Speaker5: [00:06:55] Dressed up. I think so. I was like, Should I turn on my camera?

Speaker3: [00:06:59] I was afraid I was going to be underdressed, so I'm glad that somebody dressed up always.

Speaker5: [00:07:03] This is my favorite holiday.

Harpreet: [00:07:05] Ok, that's awesome. I dressed up as a 90s grunge kid.

Speaker5: [00:07:10] Nice. I love it. I completely agree with everything, Eric said. Very much important when you're scaling. If you're in a small environment where you can directly connect to a database and race and pull your own data, then. I've done that in the past where I'm

Harpreet: [00:07:30] Very much, you know, beginning

Speaker5: [00:07:32] To end through the process. But in my newest

Harpreet: [00:07:35] Role, we have dedicated

Speaker5: [00:07:37] Data engineers and I'm finding it very fascinating and I'm very grateful for them to be able to do that work for us. And without Data, I would be.

Harpreet: [00:07:48] So yeah, they're very,

Speaker5: [00:07:49] Very

Harpreet: [00:07:49] Important.

Harpreet: [00:07:51] Absolutely love that. Matt, are you still there? Yes, you are. Matt, go for it. One point twenty one Data What?

Harpreet: [00:07:56] I love that.

Speaker4: [00:07:58] Yeah, no. Yeah, no. Like, I [00:08:00] mean, for Data, engineering is super, super important.

Harpreet: [00:08:02] I just spent all the

Speaker4: [00:08:03] Time before just playing around with just Python saying, OK, I'm just going to create a model, and that's it. But the more I've worked with Data engineers over the last three, three

Harpreet: [00:08:12] Years or so, I'm starting to find

Speaker4: [00:08:13] Out, Hey, you need to know it's garbage in, garbage out. I mean, if you don't have that Data engineer working on that

Harpreet: [00:08:20] Pipeline, you don't know if the inputs that you're

Speaker4: [00:08:23] Putting into that machine learning model are actually even worth or even accurate.

Harpreet: [00:08:27] So they

Speaker4: [00:08:29] Have to make sure that the pipelines processing, they have to make

Harpreet: [00:08:31] Sure that the

Speaker4: [00:08:33] That that the lineage is correct because not every Data that you're pulling in from one table is the original Data that Data and that table could be pulled from like three other

Harpreet: [00:08:42] Tables. And usually even as a data

Speaker4: [00:08:45] Governance analyst, you still have

Harpreet: [00:08:46] To ask them they know better than you what transformations

Speaker4: [00:08:50] Are going through and what exactly. Exactly how that's being anonymous, because there's batch thing and all this other stuff that's even being done in the pipeline.

Harpreet: [00:08:59] So it's it's

Speaker5: [00:09:01] Very

Harpreet: [00:09:01] Important that Data Engineering

Speaker4: [00:09:03] Is there and if you can, I mean,

Harpreet: [00:09:05] You need to at least have a sense

Speaker4: [00:09:06] Of it to be able to communicate with them, which is why I still do my own pipelines on the side

Harpreet: [00:09:11] So I can communicate with them

Speaker4: [00:09:13] And know what they're talking about. And they don't look at me with the question mark on their face. Yeah.

Harpreet: [00:09:18] Thanks, Matt. Yeah, definitely good to hear from a Christian as well. Before we get there, though, like that's kind of my standpoint. I feel like AIs Data scientists as people who are actually building the models and stuff like that. We're actually downstream consumers of Data. So we use Data. So it's highly, highly

Harpreet: [00:09:37] Imperative for us that

Harpreet: [00:09:39] We have some, you know, there's the real world

Harpreet: [00:09:43] And we capture information

Harpreet: [00:09:45] In the real world and the form of Data that Data goes to transformations. But like that there needs to be quality checks in place for the Data. We need to be able to ensure that it's what we expect. And not only that those Data pipelines can be robust enough to handle a wide variety of situations [00:10:00] that the engineering is. I think I came to this realization after

Harpreet: [00:10:05] Just working with

Harpreet: [00:10:06] Curated data sets for the last couple of weeks, just because now I just do machine learning all day, just building models. But Christine, let's hear from you.

Speaker3: [00:10:15] Just listening to this, and I totally agree with the importance of having that data stream and everything, but I guess I was wondering at the risk of being overly pedantic, you know, the most important part. I'm a business guy meaning

Harpreet: [00:10:28] More strategic minded. Do you think it

Speaker3: [00:10:30] Would actually be? Or maybe, maybe this is what I'm saying is where I stand is. Maybe it's actually the strategic and outcome orientation of step one of why and what levers are we pulling? Otherwise, I might have the greatest Data and I might have the greatest data

Harpreet: [00:10:43] Scientist, but it gives a crap because

Speaker3: [00:10:45] I'm solving the wrong problem.

Harpreet: [00:10:47] Yeah, I mean, but if I would have said that, I wouldn't have triggered anyone, they would see

Harpreet: [00:10:52] As an overly

Speaker3: [00:10:53] Pedantic, but maybe that's where I stand.

Harpreet: [00:10:55] It's absolutely true, though, that absolutely true Marine would love to hear from you.

Harpreet: [00:11:00] I guess that I don't know if you've

Harpreet: [00:11:01] Missed the the the question, but I was just making a statement trying to trigger people and trying to get a conversation going. That Data engineering is the most important part of Data science. I'd love to

Harpreet: [00:11:10] Get your your input

Harpreet: [00:11:12] On that. By the way, everybody in the chat, everybody watching YouTube LinkedIn. If you guys have questions, please let me know I will add you to the queue.

Speaker6: [00:11:21] Come again. What was the question, yeah, I just joined.

Harpreet: [00:11:24] Yeah, the question was that I

Harpreet: [00:11:26] Just wanted to get people's reaction to the statement that I made, which is Data engineering is the most important part of their lives.

Speaker6: [00:11:35] Um, I may have to agree with that. I, you know, I think there is a meme that has yeah. I don't even know the name in English. Like about Data Lake and I think data scientist,

Harpreet: [00:11:51] We we don't like

Speaker6: [00:11:52] We right on top of Data engineering. If that pipeline is not in place, it's not, you know, it's [00:12:00] just it's very, very hard to do and it would work and you always have to check with them, right? But again,

Harpreet: [00:12:10] But Christian said,

Speaker6: [00:12:12] I, you know, you have to have the right question to answer, like everything has to be accordingly to the outcome that you need in terms of like a product or a business or. I totally agree with that. But but again, I think Data engineering is. Yeah, I was trying to, let's say not all. We have the best. No, no. I think that I think that engineering is extremely is extremely relevant. Sometimes we take it for granted.

Harpreet: [00:12:41] Yeah, absolutely. Russell, I would love

Harpreet: [00:12:43] To hear from you and then after us go to mania for that. I'm not sure if Brussels campus frozen, no.

Speaker3: [00:12:51] You know, I think I'm getting in. Yeah. Yeah. Just can you remind me of the question I've been jumping in and out of them, if you've noticed?

Harpreet: [00:12:58] I just put a message in the chat

Speaker3: [00:13:00] Saying the Kremlin seemed to be attacking my system. So remind me if you wouldn't mind

Harpreet: [00:13:04] Know, just getting people to react to just this, my attempt at triggering triggering people with a

Harpreet: [00:13:09] Meme, just saying

Harpreet: [00:13:11] Data Engineering Engineering's the most important part of their science. I just want to get people's thoughts and opinions on that. Sure.

Harpreet: [00:13:18] Well, if we take the Data out of the

Speaker3: [00:13:21] Equation, engineering and science are two very compatible fields of very necessary disciplines for human progress. So you can't have a good engineer without science and engineering very often helps sciences, especially some of the deep sciences. You know, say I'm trying to think of a of a deep one now, but say that you know, a quantum physicist or say something like that is going to need to be able to implement some engineering principles in their science. So in that very broad spectrum, for everything in a pot to try and understand [00:14:00] the two descriptions, I'd say. Personally, me, I'm a data scientist or a data engineer, I'm a little bit of both and a little bit of a lot of other things as well. So I suppose that maybe they're sitting on the fence answer or the or the. The more. Objective and subjective is neither and both in varying degrees, depending on the job you're actually doing at the time. So some days I'll do things that I could classify more as data science. Some days there'll be more Data engineer, so. Yeah, I'm trying my best not to be triggered, if you can, you can tell

Harpreet: [00:14:41] It like the like the very politician like answer many of us hear from you.

Speaker3: [00:14:49] Hey, guys. Can you hear me? Yep. Yeah, I think my thoughts are come more from a business sense, right? Kind of like Christian picking about piggybacking off of Christian. And in a lot of organizations that I work with, they can't even have a data scientist. But what they really need is a data engineer. That's that's just frankly how what the

Harpreet: [00:15:12] Business needs, right, that they need

Speaker3: [00:15:15] To be able to access data.

Harpreet: [00:15:16] And maybe they don't

Speaker3: [00:15:17] Need the fancy data science models. Maybe what they need is. Well, I just need this view and this visual right.

Harpreet: [00:15:23] So they need

Speaker3: [00:15:25] To get all these sources together. So I think we're trying this is a data science podcast, but it really goes to what what does the business need? What are we trying to solve a problem or business problem? Or are we trying to solve? And sometimes there may not be a data science problem?

Harpreet: [00:15:41] So that's why

Speaker3: [00:15:43] I think data engineering in some sense. It's kind of like the foundation for everything, especially if you're trying to go into analytics and as everybody else was mentioning, data science builds on top of that.

Harpreet: [00:15:57] Yeah, yeah. I'm kind [00:16:00] of liberal with my use of the term data science. I considered like data science to be like an umbrella. So there's like the traditional data scientist, Data engineer, analytics engineer by people, you know, data analyst. All these people kind of fall into that data science umbrella. But yeah, I'm ready to take some questions. If you guys got any questions, please do. Let me know I'd be happy to take some of these questions on. Oh do by the Facebook changed name to Metta, huh? That's a that's I mean, you know, Ron Artest should say something I don't know if y'all know who

Harpreet: [00:16:33] Ron Artest is basketball player

Harpreet: [00:16:35] Who changed his name in the mid 2000s to Metta World

Harpreet: [00:16:40] Peace.

Harpreet: [00:16:41] I think he has the the rights to that. But yeah,

Harpreet: [00:16:45] Anyway, is that what the

Harpreet: [00:16:46] Companies are no longer saying? But it will be, I guess,

Harpreet: [00:16:48] Mang or Monga? Hmm, interesting.

Harpreet: [00:16:52] Marina, go, let's hear from you.

Speaker6: [00:16:57] I know I have a question about these. So what kind of visualization tool like is everybody using? And probably if you will have to learn one or you want to use one outside work because you know that maybe a different, different one? Which one you will recommend like Locker? I don't know, like Tableau or yeah, that's that's my question.

Harpreet: [00:17:26] I'd love to hear from

Harpreet: [00:17:28] Anyone who has expertize on this, I don't do a lot of visualization visualizations, just a map plot lib, but go for a Christian.

Speaker3: [00:17:34] Yeah, so that's how I'm person. That's how I got

Harpreet: [00:17:36] Started in the Data world is I got introduced

Speaker3: [00:17:38] To Tableau and Power BI and all that kind of stuff I found. If you can just like that, anything is you kind of learn the lay of the software, right? Tableaus, super simple and super applicable. And with Tableau Prep also like it helps you to do some basic, repeatable cleaning, right? And so that's what we deployed [00:18:00] when at a larger company I used to work for, we kind of spun up an analysis group across functions. We had a lot of data and information that we were able to acquire ourselves, but didn't have a great way to present it to folks. Tableau is great because you drag and drop it in there. You create all these views interactive and you can publish it to server as well and so stakeholders can interact with it now. And so I have a really, really good experience with with Tableau. There's some things that that Power BI does. Tableau doesn't, but I'm still a Tableau fan. So if anybody asks and they're looking to get started,

Harpreet: [00:18:29] I recommend Tableau myself.

Harpreet: [00:18:31] Matt Blousy, go for it.

Harpreet: [00:18:36] I mean, if it's just for like building out your

Speaker4: [00:18:38] Own visualizations on the side,

Harpreet: [00:18:39] Tableau, I agree with Christian is really

Speaker4: [00:18:41] Very, very useful for that.

Harpreet: [00:18:44] If when you're

Speaker5: [00:18:45] If you're working like with something more like

Speaker4: [00:18:46] Azure or you have an azure environment or like a server like their power BI, I found, is a lot more easy to configure

Harpreet: [00:18:53] To get insights for that

Speaker4: [00:18:56] And to connect to the Data than Tableau is.

Harpreet: [00:19:01] Anybody else have any insights

Harpreet: [00:19:02] On this, I want Ben Taylors

Harpreet: [00:19:04] In the House, by the way, Ben Taylor, what's going on and I

Harpreet: [00:19:05] Could see like that with

Harpreet: [00:19:08] It at the beginning there are two because so we call it in Canada. Russell, go for it.

Speaker3: [00:19:14] Yeah. So I just wrote something very quickly in the chat, they're saying trying to decide the best visualization or AI tool. It's like trying to decide the best dessert. Your favorite dessert. Michelin starred chef is going to be able to create something fantastic that not very many people at home could do.

Harpreet: [00:19:31] But your average

Speaker3: [00:19:31] Cook at home might have a really good cheesecake recipe or, you know, a chocolate cake recipe that many people can do. And I think it's kind of like that if you are a Michelin starred chef level of Data, let's say broad Data, not just Data science engineering, but any kind of analytical approach to Data. You're very possibly going to have the skills to use, you know, plot all of those kinds of things. But the more [00:20:00] approachable ones such as Tableau, Power, BI, etc. And great for I don't want to use the word novices, but you know, the lower level echelons of that Data expertize, especially power bi. It's part of the Office 365 environment, so it connects with so many things you can you can run it, stand alone, you can run it on an enterprise level, you can connect it to AIs or SQL, etc. So there's a lot of room for expansion for that. However, it's tied into the Microsoft environment. Tableau standalone

Harpreet: [00:20:36] Also

Speaker3: [00:20:36] Very good in its own right, a little less flexible than Power BI, so it kind of depends on the environment in which you're using it. The skill base of the users and what I say, the users.

Harpreet: [00:20:49] I would draw

Speaker3: [00:20:50] Distinctions about the user developers and the user consumers. So both of those you will publish something, the users

Harpreet: [00:20:58] Will come in, they will interact

Speaker3: [00:20:59] With it via the, you know, the slices of filters, 17 pages, et cetera, and their slightly different usage characteristics between the two. And then your developers as well will create and maintain all of the stuff that you publish. So there's a, I'd say, kind of a curve scale, your entry level things and as you say, powered by Tableau, there are others click and a whole host of others that I can't remember off the top of my head. But those are a great entry level tool. But don't restrict yourself to using those simply because you start with those. They should not restrict and allow you to grow and build upon those as part of a wider visualization structure.

Harpreet: [00:21:46] Thank you very much, Russell. Anybody else had any suggestions, Marina, did that

Harpreet: [00:21:50] Help at all? Like, are

Harpreet: [00:21:51] Any of these things free? Like, I know Tableau has a free public tier, but I

Harpreet: [00:21:54] Don't know about the rest of

Harpreet: [00:21:55] These things. Like, Is there anything else?

Speaker3: [00:21:59] Be AIs free, [00:22:00] but I find it. It's a little bit of a frustrating licensing model. So it's free for everybody. You can download the desktop and you can do whatever you want. However, you cannot see the product of anybody else unless you pay for a paid license. And the first paid license, I think, is a pro license. I think it's 9.99 per user per month. So you can't see others work until you have a pro license. It's kind of a reverse pay structure, and it's almost like I had this conversation with a few other people. It's a bit like a kind of a drug pusher, you know, give you the product for free that you play around with it, get deep into it, really enjoy it. Then when you want to share your work, you have to actually ask other people to pay for a license so they can see your work and vice versa.

Harpreet: [00:22:42] So there's benefits to it.

Speaker3: [00:22:44] But there's also distinct restrictions to it. And tablet, I think, is the other way around. So there's different license tiers to it. There's kind of a basic user which is low in the midst of what powered by low licenses, and there's a developer license and the server hosting modules that also change much as they do with LBI.

Harpreet: [00:23:06] Grew up around a lot of drug dealers, so I can say that strategy does actually work. Let's see. Gina had a question that came in before Christian's. Let's go to Gina's question, then we'll go to Christian's question. Gina, you had a question about deploying models. I think Ben Taylor or Jodi

Harpreet: [00:23:22] Would love to hear you

Harpreet: [00:23:23] Guys perspective on this. So you're out for the question here.

Speaker5: [00:23:27] Great. Thanks. Yeah. Hi, everyone. Good to see everybody. So my question is around. And for those of you who haven't interacted with on the happy hour yet, I'm I've been through a Data science bootcamp. I also have many, many

Harpreet: [00:23:46] Years of professional

Speaker5: [00:23:47] Experience, including some analytics, financial modeling, investment analysis

Harpreet: [00:23:53] And a whole bunch of other

Speaker5: [00:23:54] Things. So I'm coming at it with a career pivot,

Harpreet: [00:23:59] And [00:24:00] I'm

Speaker5: [00:24:00] Continuing to build my skills after the bootcamp. And so I'm aware of some tools to deploy models or even just little software scripts in online.

Harpreet: [00:24:12] And so like, I'm an

Speaker5: [00:24:14] Avid cyclist, but particularly on Zwift indoors, and I joined actually, I joined a racing team, which is kind of funny. I'm going for last place because I'm racing above my level, but somebody came up with the idea. They want to be able to post in discord.

Harpreet: [00:24:30] They want to be able to post their

Speaker5: [00:24:32] Kind of workout graphs from the workout, but they want it to look like a Zwift graph. Just so just ease of comparison. You know, you see it at a glance, you know where people were in their workout zones. And so somebody started some code and I'm like, Ooh, I'm going to take this, and I built on it and added some stuff, and now I'm thinking of putting it up in streamlined. So that's a very, very simple example. That's not a data science thing per say, but one of the things that didn't really go into in the bootcamp and I really wish they had was, you know, deploying models with, you know, flask or or breast. And like, I'm already kind of out of my element, so I really like to get your guys's input. I know for a lot of you, this is like probably super basic entry level stuff, but I'd really appreciate your thoughts because as I continue to build on my portfolio, this is an area where, you know, it'd be great to understand some of the good, some of the better places to go with it.

Harpreet: [00:25:36] Yeah, definitely, I mean, you'd be surprised,

Harpreet: [00:25:38] Like in a

Harpreet: [00:25:40] Lot of companies, deploying is not

Harpreet: [00:25:41] Fancy is just, you know, sometimes

Harpreet: [00:25:43] Saving solutions off as a CSV, pushing them into a database. Some people get sophisticated, have a web API if they want to be more advanced. But Ben Taylor, let's go to you.

Speaker7: [00:25:53] Yeah, if it's a personal project, flask

Harpreet: [00:25:55] Is great, so Gina,

Speaker7: [00:25:56] If you look in a flask, I think you will be delighted how simple it is to start. [00:26:00] I'd recommend looking. I think the classics like the chat server, so you have like a chat server example, then you'll realize like, Oh, so when I ask a question, when I send something in, I'm going to have a model loaded in memory and I'll kick back a prediction and I think you'll find lots of flask flasks. You learn examples, but I think if you look into it, you'll be delighted that

Harpreet: [00:26:19] Like, Oh, this is super easy

Speaker7: [00:26:21] For anything that is being consumed in production or company.

Harpreet: [00:26:27] You know, there's

Speaker7: [00:26:28] Issues on scalability, sleighs, there's consequences. And so the the list of consequences quickly grows. It's not just about the model going down and scalability issues, it's also stuff like feature drafts. So nice, higher view. We had a feature drift excursion that impacted the customer, which surprised us because the models are static. So model static customers happy first month, second month, 3rd month. Everything is great and then settling within twenty four hours a terra on fire. The model's not working at all and the root cause. It doesn't matter how good your your data governance is, the root cause is a vendor

Harpreet: [00:27:03] That doesn't have as good as data governance as

Speaker7: [00:27:05] You, especially when you get to these really wide models where you're consuming stuff from other people outside of your circle. They had shifted a threshold without telling us thinking it was minor

Harpreet: [00:27:14] And it had been

Speaker7: [00:27:15] It had impacted something in the model. So, yeah, so when it comes to production where things are consumed, I think there's a very long list of problems. But I'm curious like, I'm sure there's people on this call that have started with the put a flask wrapper API around your machine learning model and then respond to scalability like you can throw machine learning models on lambdas on Amazon. You'll get really good

Harpreet: [00:27:38] Scalability, and then

Speaker7: [00:27:39] Maybe something I'm more naive to is you do have sage maker and different things that can scale models. But I think it's still a little bit more. I have a bias because I work at Data robot. I'd say it's still more like the hobbyist level when it comes to like, Oh, you get fired for a prediction gone bad, then I would lean towards my employer for stuff like that. Or I'm curious, Joe, are we competitors?

Speaker8: [00:28:01] I'm [00:28:00] sure you'll be fine. Then you can send me money after joking.

Speaker7: [00:28:07] Yes. Here it comes.

Speaker8: [00:28:11] Yeah, I think that those are all good points, really. And I think for where you are

Harpreet: [00:28:14] Right now

Speaker8: [00:28:16] Focusing on, I think, showing that, you know, how to deploy a model is the most important thing. The other things I would say you should be should know about, too is just how does rest work? You know, like, I think that's an understated skill and this sort of starts getting into software engineering. But when you're deploying stuff to production, you can ignore software engineering and so know about like also other things like serializing and sterilizing Data. That's pretty key and just a basic concepts of like, you know, rest like, you know, get put post delete requests and that kind of stuff. I think they're just awesome things to know in general. Like it'd be one thing to know how to, I think, going through a tutorial and deploying flask. But then like, why does it do what it does? I think those yeah, where you are, I think those are the more important questions I would say like you should be looking at because like, as Ben points out, once you get into production, there's

Harpreet: [00:29:09] No shortage of

Speaker8: [00:29:11] Myriad ways that things can go wrong. And also knowing like basic stuff like this, my eight, if my API is broken like where I tend to be understated quite a bit, that's what I would

Harpreet: [00:29:24] Personally also look. But I think what Ben

Speaker8: [00:29:26] Suggested is great and just go use Data robot for everything.

Speaker5: [00:29:31] So, so. So is there would you say there's kind of a hierarchy you kind of alluded to it, right? I mean,

Harpreet: [00:29:39] You know, right now I'm talking

Speaker5: [00:29:41] Streamlined and a a guy who is also in my program use is using streamlined on a very basic,

Harpreet: [00:29:49] You know, predict the

Speaker5: [00:29:51] Grade level of elementary school students, you know, predict their grade

Harpreet: [00:29:55] Level based on

Speaker5: [00:29:57] The essay. And literally, it's like you upload the text [00:30:00] in there and then it makes those predictions. So but it sounds like, I mean, you have stream lit, you have flask. I don't know where rest falls into that, because that's

Speaker8: [00:30:11] What I'm saying. Like, learn, learn, note rest is because it's not, I think, extreme and flat sort of obfuscated. A lot of these details, like rest is really a paradigm that Santa

Harpreet: [00:30:19] Represents representative

Speaker8: [00:30:20] Representational state transfer, sorry talking for days now. And really, that's

Harpreet: [00:30:25] A paradigm, you know,

Speaker8: [00:30:27] Of of it's a way of not asserting Data, but it's a way it's a request and response model of Data. Right. And so that's what I'm saying. Like Flask is just it's it's an abstraction built on top of actually, flask itself is a web server and a web framework. There is actually flask rest for the API part, right? And then there's also new APIs like Fast API, which I think is dope. But if you don't really know like rest and kind of how the request response model works in web servers like that's where I would say, like, start there, because once you know that, then it's kind of like any API is an API.

Speaker5: [00:31:02] So and then where does it go from there? I mean, you mentioned Data robot, which is up here somewhere. I mean, you mentioned Lambda on the. And you said, even with that, that's not necessarily at a level you would want to be when your job's kind of on the line really day to day.

Speaker7: [00:31:22] There are so many options.

Harpreet: [00:31:24] A lot of

Speaker8: [00:31:24] Options. Yeah. The problem, right? I mean, I think most of the discussion is just like trying to like cut that are like, what is a tool versus what's a paradigm? Yeah. I kind of has this notion of serverless. Now you've got to figure out what that is.

Speaker7: [00:31:39] We did a webinar this week where we're showing predictions in job

Harpreet: [00:31:43] In Snowflake, so like natively

Speaker7: [00:31:45] In Snowflake, you can upload Java user defined functions

Harpreet: [00:31:48] And then right there in SQL, you're doing your machine learning models,

Speaker7: [00:31:51] Which is pretty awesome. So then there's no server Snowflake took care of it, right?

Speaker5: [00:31:57] Just I was speaking with someone this [00:32:00] week who has actually been at a company for like over 10 years they've been doing modeling. The company has some patents actually in, I think, some machine learning stuff going way back and he was talking about they use Google BigQuery, I think,

Harpreet: [00:32:18] And they tried

Speaker5: [00:32:19] Snowflake for a while. And then they actually I think it just didn't work very well for their use cases. So it seems like now

Harpreet: [00:32:26] You're talking

Speaker5: [00:32:27] Google BigQuery. I mean, that's getting into those higher, you know, kind of more heavy duty approaches to deploying models. He's a data engineer and has been for many years, so he knows his stuff.

Harpreet: [00:32:42] Yeah, that's

Speaker8: [00:32:44] Well, that's another one, that's a data warehouse that's serverless, and so you can see it's kind of like turtles all the way down and. Oh, and the other thing I forgot to mention, too, if you're if you're dealing with an API, then you know how networking works. If you're going to do it in production, you better damn well know networking security works. You're not going to have like some unencrypted

Harpreet: [00:33:02] Thing of

Speaker8: [00:33:03] That authorization. I mean, you'd be

Harpreet: [00:33:04] Crazy and you get fired.

Speaker8: [00:33:06] So you got. There's a lot to know. So I say, like, you know, it's interesting, and that's where I think Ben is sort of alluding to as well with these tools that sort of allow you to work in production because if you were to try and do all this stuff by hand, this is what I was seeing. A lot of data scientists are like, Oh yeah, I'll just do this model and production and flask, and it's like, cool, you know what that actually requires in production. It's not just like firing up in your laptop and on local host. It's there's a lot to it.

Speaker5: [00:33:32] Thank you for mentioning Joe

Harpreet: [00:33:34] About rest

Speaker5: [00:33:35] And making that distinction that it's a paradigm and really understanding rest, it sounds like we'll

Harpreet: [00:33:41] Help a lot just conceptually and

Speaker5: [00:33:45] More tactically understand how stuff is actually the what is happening under the hood. If that's accurate to say in deployment and learning, some of these things kind of a fundamental level. Look. [00:34:00]

Harpreet: [00:34:01] Thanks. Right. Great question, great discussion here. Let's go to Kristen's question. Yeah. Christian, go for it.

Speaker3: [00:34:11] Yeah, sure.

Harpreet: [00:34:13] You know,

Speaker3: [00:34:14] For the last 10 years, it seems like it's been a lot of questions about resources for learning skills, and that's always going to be important to stay sharp.

Harpreet: [00:34:19] But I'm just wondering for the folks

Speaker3: [00:34:21] Here like, you know, what are the best resources that you might recommend that you've come across for Data scientists specifically to get more effective at things like identifying how high value use cases and informing and influencing stakeholders?

Harpreet: [00:34:35] Within the Shysters newsletter, that's for sure, his newsletter is amazing.

Harpreet: [00:34:41] Mostly newsletters.

Harpreet: [00:34:42] That's straight. There's this other guy that I found on. Shooter's name is Santiago. I think if you type in like Santiago Machine Learning or Santiago de Data scientists who might come up on on Twitter. His stuff is pretty good as well. But yeah, I'd love to hear from you. I mean, for. What's anyone really like? I think Eric dropped off, but Monica, how do you stay sharp? Matt, how do you stay sharp?

Speaker5: [00:35:13] Um. For more strategic level stuff, I tend to look into the methodologies that you're using. So. Within cybersecurity, there is that Nest eight hundred and fifty three type stuff, and just look at that high level. I don't. Yeah, I can't think of anything beyond like those high level methodologies.

Harpreet: [00:35:43] But let's hear from a band or Joe on this because he has been in the game for four for a while doing this stuff. O.g. days, how are you guys keeping on top of things?

Speaker7: [00:35:53] Christian, you're talking about influencing key stakeholders

Speaker3: [00:35:56] For your sure. Yeah, I'm from a data scientist perspective, especially like, [00:36:00] you know, you're always talking about, you want to be outcome oriented, you want to hit the big wins, especially, let's say you're early on in the data science or you want the quick wins. How do you identify the high value use cases? And then kind of along with

Harpreet: [00:36:13] That is like now you've got to influence

Speaker3: [00:36:14] People that these are the right things to do.

Speaker7: [00:36:16] Yeah, no. I love this question because I really sucked at this because, you know, we all love data science. We're so excited. And I remember having like an hour long meeting with an executive and showing them the three things I tried. And now the third one is working so well and they don't understand any of it. And I'm thinking like the whole time they're wondering, like, how much are we paying you? Should we fire you? So the best way to do it is there's some good exploratory questions, so it's good for you to know what are your top KPIs this, you know, this quarter? Because how are they getting bonus out? Like if you can help them next quarter, they're really going to like you. So what are your top KPIs? And then if you want to be more exploratory, you can say, what are the numbers in your department where small change in that number makes you really excited? And then they're going to talk about, Oh, we've got loss, they'll bring up some numbers and those could be opportunities for A.I. improvements. And then the other thing I like to say is if where where's your growth bottleneck of human capital? So if I could gift you a thousand humans, they're experts.

Harpreet: [00:37:12] It's only one job family. What are they in? They're free. They're free for a year. Where are you going to put them in your department?

Speaker7: [00:37:17] And so that allows you to kind of shine a light on potential problems. But the more you start working backwards,

Harpreet: [00:37:23] The best thing you can show

Speaker7: [00:37:24] And executive is a number with a dollar sign. And that confuses Data scientists because we're so analytical, we're so fact-based. But the reality is you can go through a utility function and come up with all these assumptions. So if you have a predictive model and you do some kind of cut or something, you can back that into, I'm estimating I'll save you $10 million a year. And if I disagree with you and say, I think you're full of it now, you can go through your assumptions and we can have a discussion. So the more you work backwards, never use any jargon, never show your work, just show the outcomes that you'll do really well.

Speaker8: [00:37:56] I did not know that originally. I've been still [00:38:00] very successful, so

Speaker7: [00:38:02] I've had my ass kicked enough. But you really do win on the quarter.

Harpreet: [00:38:06] So if you can guarantee, like

Speaker7: [00:38:08] Everyone you interact with, how are you going to get them promoted next quarter? With that mindset, you will be working on the right problems. Obviously, you want long term strategy, but you can't do long term strategy. If you don't have short term wins. First, get the short term wins, then you can be more expansive.

Harpreet: [00:38:24] Thank you very much, Ben. Love to hear from you on this, and after that, we'll go to Gina.

Speaker8: [00:38:30] I mean, it's. This is very much a sales question, right? I think that's the most things end up just becoming a sales and marketing at the end of the day. So. Yeah, I mean, Ben's right start with kind of the problem at hand, like, you know, it was on

Harpreet: [00:38:49] Three or four sales calls today, right?

Speaker8: [00:38:52] I don't think once it ever use jargon, we're talking about stakeholders, some technical, some not. And the thing I was focused on was like, What are you trying to do and how can I help you? That's it, right? At the end of the day, you know, when I ask a really wealthy friend how he how you make a lot of money, he said, Well, you sell what people want to buy. It's easy. It's that easy in that hard, but really, that's no different when you're trying to sell a project internally or something,

Harpreet: [00:39:24] Right, you just figure out what

Speaker8: [00:39:25] Motivates the person and what they value and sell them that you don't need to like. And in fact, it almost talking too much works against you in a lot of ways that same guy is holding when, like when they're about to buy to shut up and take the sale. Stop talking. So that's the other thing, because to say is scientists will all be tempted to talk a lot. Like I am right now, but.

Harpreet: [00:39:50] 48 laws of power

Harpreet: [00:39:51] Law, number four, I always say

Harpreet: [00:39:53] Less than necessary. When you are trying to impress people with words, the more you say, more common you appear [00:40:00] and less in control. Gina. Go for it.

Speaker5: [00:40:04] Ouch. Now I'm going to say something right after that. Oh my God, it's so true, and I worked on it for years and I'm still working incentives.

Harpreet: [00:40:14] It's all about

Speaker5: [00:40:15] The incentives that your executives, your managers are facing. And that's not to say they'll always be rational. I've heard of situations where the data scientist does this analysis. They try their best to explain it, and it goes against the executive's intuition. I won't go into the specific case, but. You know, it's it can be crazy sometimes, so I mean, you can't there's only so much you can do. But like Joe says, like Ben says, and this is where

Harpreet: [00:40:46] Having lots of

Speaker5: [00:40:47] Years of experience prior, you know, in a range of environments, including consulting, it's the bottom line. Why do they care? Why should they care? What, what matters to them? They don't care about fancy statistics or cool models.

Harpreet: [00:41:04] I mean,

Speaker5: [00:41:04] If they're not technical, they're not going to care about any of that as long as they know

Harpreet: [00:41:08] That there's something behind it and

Speaker5: [00:41:10] You're not just making crap up. It's really important. And so years ago,

Harpreet: [00:41:17] Before I was doing data science stuff,

Speaker5: [00:41:19] I was working in a role that was worldwide team and a great big company. No, pal. So you have to try to earn your keep somehow. And we were working with two different groups trying to convince

Harpreet: [00:41:36] The technology

Speaker5: [00:41:38] Solutions group the people who would actually go out and repair PCs and repair servers and this and that to to push some of that out. Let more of that be handled by channel partners, but they didn't want to do that. They kind of wanted to own all the customers. And this is a problem from a market share perspective, especially in certain countries, [00:42:00] because they're getting their lunch, you know, they're getting their butts handed to them in certain countries by other big hardware manufacturers at the time. And so what we did was and I came at it with a, I'm agnostic, you know, I'm going to go out and get information for people. And we talked to a lot of salespeople in different regions and they kept coming back saying, Yeah, you know, this other big storage company, you know, kicked our butts. I lost $6 million sale here.

Harpreet: [00:42:29] I lost 10

Speaker5: [00:42:30] Million there because for the small and medium business customers, they were fine with letting channel partners handle it. And our company wasn't. And this was extremely contentious.

Harpreet: [00:42:42] And yet we

Speaker5: [00:42:42] Managed to with the analysis I did with the help of some others, we managed to convince the senior management in that

Harpreet: [00:42:50] Group that didn't want to

Speaker5: [00:42:51] Give up that channel managed to convince them to at least give it a try.

Harpreet: [00:42:57] And that was, you know, a big

Speaker5: [00:43:00] Win

Harpreet: [00:43:01] For our group because

Speaker5: [00:43:03] We needed to have credibility and we needed to show that, you know, even though we were kind of more aligned on the hardware side, we needed to be able to show that this is legit and back it up

Harpreet: [00:43:15] With analysis. But at the bottom

Speaker5: [00:43:17] Line was we could be making $40 million more annually in basically, we're growing the whole pie, so we're making more sales even on the technology support side by giving up some of the stuff to channel partners. And when you can make a case

Harpreet: [00:43:35] Like that, it's pretty

Speaker5: [00:43:36] Hard to for anybody to argue with it, even when there are fiefdoms and

Harpreet: [00:43:41] Territorial kind of

Speaker5: [00:43:43] Turf wars going on. So that's the other thought I would convey on that is just also understanding those dynamics and establishing trust with the different groups. So you have credibility. They don't think, Oh yeah, well, you're just coming from this group and you're just trying to, you [00:44:00] know, and in big companies, you can have these these tensions sometimes. So I probably talked too much there again, but I hope.

Harpreet: [00:44:09] Gina, thank you so much.

Harpreet: [00:44:10] Appreciate your perspective. Thank you so much.

Speaker7: [00:44:14] So this gets around real quick. Listening to Gina, I think one of the things I'm reminded of, that's why it's important to have a shotgun approach, because if you bet all your eggs on one model and there's some intuitive hiccup or pushback or politics you weren't expecting,

Harpreet: [00:44:31] Have it have three to five problems

Speaker7: [00:44:33] Work with the subject matter. Experts have them right. The dollar signs next to the problems before you start because you can get buy in before you even start. I think it's funny because you say, What is this worth and you have shoulder shrugs, then say you can put a zero one dollar sign or up to five dollar signs. And then if they tell that to you, then you're kind of working on the right problem, and

Harpreet: [00:44:53] That might save you from

Speaker7: [00:44:55] Stubbing the toe on something.

Harpreet: [00:45:00] Jason, hopefully that was helpful. That's a lot of good insight that thanks so much, guys. A let's let's keep moving, we got a question coming in. Greg, Kogelo is chilling in the chat. Before we get to Greg's question, though, shout out to Antonio Sinofsky. He is watching on LinkedIn Antonio. They just had him. His wife just had a baby a few weeks ago. Congratulations to

Harpreet: [00:45:22] My friend.

Harpreet: [00:45:23] Hopefully you got the Snoopy. That thing, man. You going to need that the key to sleep in? But congratulations to you, man, can be a hell of a ride. So great question coming in from Sir Ben. And he wants to know. I saw this post that you posted just as soon as I opened up LinkedIn, it was. What do you think about Ellen's idea of a Texas

Harpreet: [00:45:45] Tech and science

Harpreet: [00:45:47] University?

Speaker7: [00:45:48] So I'm a little embarrassed about this because it's a joke, so it spells tits, but on a huge Ellen fan. So I felt I fell for the joke. I feel like anything Ellen touches gets disrupted, and so [00:46:00] unfortunately it is a joke and it's inappropriate. But if Ellen did say I'm going to make a university, I would fight like hell to get my kids

Harpreet: [00:46:07] To go there. And I think all

Speaker7: [00:46:09] The other institutions in institutions would be in huge trouble because it takes five years on average for an

Harpreet: [00:46:15] Institution like your traditional

Speaker7: [00:46:17] Institution to incorporate new curriculum

Harpreet: [00:46:19] And for everyone on this Zoom call.

Speaker7: [00:46:23] How does that set them up for A.I. in Data science like it's, you know, they're like failing before they start, and there's a few that are OK, and that's why these bootcamps

Harpreet: [00:46:31] Are so appealing because they're much

Speaker7: [00:46:33] More nimble. So. Yeah, sorry, I kind of I spread I spread that joke even further, you know,

Speaker3: [00:46:41] Then I wouldn't say that I wouldn't even put it past him for not to be. I mean, look at the way that he's named his models for the

Harpreet: [00:46:47] Tesla, but they're real

Speaker3: [00:46:48] Cars, so he may very well put a university out there with the acronym kids.

Harpreet: [00:46:54] We ought to explain that to me.

Harpreet: [00:46:56] What's his?

Speaker3: [00:46:59] So the Model S, the Model

Harpreet: [00:47:00] Three, which he wanted

Speaker3: [00:47:02] To be the Model E, the Model X. Right? I mean, he really did it. That's what he does. So he's, I wouldn't put it past him to make a university and make that acronym. And I agree with Ben starts like that. I've been thinking for I've been in college for way too long, just started another master's program and some of them are great. But there's a lot of friction, a lot of pain, a lot of just dumb stuff happening in the university system ripe for disruption. And I would totally this has been with saying, I totally agree.

Harpreet: [00:47:30] What a guy, man.

Harpreet: [00:47:31] I love that guy, though, seriously,

Harpreet: [00:47:34] Greg, there you go. It's a joke.

Harpreet: [00:47:36] So hopefully your question wasn't wasn't serious. Then you weren't just making fun of that, Ben. So anybody else has a question, whether you're on LinkedIn, whether you're in the chat room, let me know. I just want to, you know, ask a question, mostly for four

Harpreet: [00:47:49] For Joe, because I know

Harpreet: [00:47:51] I've heard you talk about this before the switch from ETL to E-ELT, the use of something like, you know, DVT like break this [00:48:00] down for a for for data scientists and analytics engineers out there. And you know, what's the shift about? Is it a cultural shift? Is it just,

Harpreet: [00:48:09] You know, we realize

Harpreet: [00:48:10] That it's easier to do this? Yeah, you

Speaker8: [00:48:13] Know. Um, yeah, I mean, so the difference really is there's etal, which is extract, transform and load, right? And then there's ELT, which extract, load and transform EDL for the longest

Harpreet: [00:48:25] Time has been the predominant

Speaker8: [00:48:27] Paradigm is typically used in data warehousing, for example, removing Data into Data lakes. The idea is with ETL, you can extract data from a source system. You're going to transform it in flight and then you can land it, maybe into a database

Harpreet: [00:48:38] Table in a destination. Right?

Speaker8: [00:48:40] Subtract. Transform the data, then load it. And the reason why this came about, really, it started because. I wouldn't say it was adopted because you

Harpreet: [00:48:52] Had limitations on hardware back in the day, this is

Speaker8: [00:48:55] Predominantly a paradigm used an on premise systems where the white servers, databases, these are expensive and you don't have a lot of resources to throw at it, right? So you have to typically use the system to the transformation. You know, it's going to get bogged down right and you don't want to interrupt the downstream database from queries, nor

Harpreet: [00:49:14] The upstream system from

Speaker8: [00:49:16] What you're getting

Harpreet: [00:49:16] Data.

Speaker8: [00:49:17] But, you know, you kind of circle around. The advent of cloud data warehouses starting with redshift, I think the realization was, well, you could just throw your data into you, extract the data from a source system and then loaded into a cloud data warehouse, for example. A staging area, just that raw data is there, you could use the seemingly infinite

Harpreet: [00:49:38] Compute power,

Speaker8: [00:49:40] You know, of the of the cloud data warehouse and transform the data there and just load that into a new table within that data warehouse. I think that was really the fundamental shift is what happened was cloud data warehousing and along with that, to some extent, cloud data leaks, although I think that. It's more of a separate discussion, but we'll stick with data warehousing pronounced [00:50:00] that really was was the genesis of, I think, the L'T movement and since then it's. But I would say this is becoming more of the dominant paradigm we're seeing etal is becoming I would see less of something you would do as a net new. Way of getting Data into your Data LinkedIn warehouse, so

Harpreet: [00:50:22] That makes sense. But the revamped version of it?

Harpreet: [00:50:25] No thank you. I appreciate that. Yeah, definitely enough to check. I think you did like a talk or discussion about this. That was a dedicated that you did this talk.

Speaker8: [00:50:34] Yeah, the expo got a ten minute lightning talk.

Harpreet: [00:50:37] Yeah, yeah. Go ahead and check that

Harpreet: [00:50:39] Out as well. Thank you very much. Yeah, it's been.

Harpreet: [00:50:41] Yeah, I'm always I just want to get into all parts of Data science like, you know, I just want to know a little bit of

Harpreet: [00:50:47] Everything just to be well versed

Harpreet: [00:50:48] As possible. So Data engineering is kind of been the thing of, you know, kind of

Harpreet: [00:50:54] Poke around it and

Harpreet: [00:50:56] Learn more about. All right, let's see if there's any other questions. Shout out to Ken McCabe found his way into the room when I realized that I had put the link a faulty link in the

Harpreet: [00:51:09] In the chat, but I'm

Harpreet: [00:51:10] Glad to hear. Do you have any questions? Let me know. I appreciate it. Yeah, no problem. If you have any questions, let us know, Matt. Youtube or LinkedIn any questions? Please do. Let us know. Right. No questions coming in, I just. We won't take this.

Harpreet: [00:51:35] Let's see. All right.

Harpreet: [00:51:36] This is a host in crisis, just trying to think of something to come up

Harpreet: [00:51:39] To with two, I discuss

Harpreet: [00:51:41] Benazir unmetered. Please save me.

Harpreet: [00:51:44] What is next

Speaker7: [00:51:45] After deep learning?

Harpreet: [00:51:47] What is next after deep learning?

Speaker7: [00:51:50] Can deep learning take us where we need to go? Or does it need a reinvent? We need to burn it down on the ground and start over. In the first of anyone,

Harpreet: [00:52:00] Do [00:52:00] the physics,

Speaker8: [00:52:01] What prompts you to ask

Harpreet: [00:52:02] That question? Yeah, yeah.

Speaker7: [00:52:04] Because I'm obsessed about the singularity in my kitchen.

Harpreet: [00:52:07] Yeah.

Harpreet: [00:52:09] I mean, look,

Harpreet: [00:52:10] I think it's going to be the analogy between how far in Newton's physics has gotten us now and then, you know, quantum after that. Who knows what's after quantum, right? I was listening, you know? At least partway this interview that Lex Friedman released with Stephen Wolfram. And there's talking about all this crazy cutting edge stuff with complexity, theory and and and consciousness and all that stuff and I mean, we have no idea what's going to happen next, right?

Harpreet: [00:52:42] Like we're building the future, literally.

Harpreet: [00:52:45] This is probably the only time in human history we're building. We are literally building the airplane while we're flying it. You know, I mean, yeah, that's kind of the analogy, I think

Harpreet: [00:52:54] Is like, OK, well, you know, everybody

Harpreet: [00:52:56] Thought Newton's physics was it, and then quantum mechanics came along and who knows what's after quantum mechanics?

Speaker7: [00:53:06] I think the general trend is like faster, easier, like because people are so excited about semi supervised learning, which I think is a little silly because it's like so obvious, like, oh, you want to label some images, but you don't want to label a million of them. Maybe you should train a crappy model when you have one labeled like, it's funny because people say it's like it's so profound. It's like no dummy. Like, of course you should do that.

Harpreet: [00:53:28] But I think when it comes to

Speaker7: [00:53:30] Better and faster, you can imagine some really, really brilliant, really, really good, unsupervised deep learning approaches where the moment you label like five or 10 images,

Harpreet: [00:53:39] It's like

Speaker7: [00:53:40] Just like taking off or it's done like some clustering or something. I get really

Harpreet: [00:53:46] Excited about this

Speaker7: [00:53:47] Concept of A.I. AI that survives to please you and by please you. I mean, actual insights. So wrinkle

Harpreet: [00:53:55] Their camera app

Speaker7: [00:53:56] Is awful like it's you can't turn it on because it just blasts your [00:54:00] phone all day. That's not actual insight. So I think A.I. of the future, everything is actionable, regardless of

Harpreet: [00:54:06] What class of deep learning you think will take us there. I mean, it sounds

Harpreet: [00:54:11] Most likely

Harpreet: [00:54:12] Deep reinforcement learning. Some flavor of that will probably be huge in in in the future.

Speaker7: [00:54:20] Yeah, I think. It's I think there because there's been some excitement around novelty based systems or focus based learning like and I apologize too, because I've actually I used to drown in this stuff like I would like choke on it. And this is why this was my life for three years and it's

Harpreet: [00:54:39] Been I went a

Speaker7: [00:54:40] Full year at Data robot without programing, and I'm programing again coming back into the fold. I think that you are going to have you're going to have deep learning networks that can be much more impressive training from zero. So GPT three in these big nets that train with more data than any human can ever imagine, you're going to have some deep learning examples that shock you with their ability to learn with limited observation. And when I think of those systems, I think of systems that are really good at novelty because because right now deep learning is not. If I throw like an outlier into the training set, it's not going to do very well when it sees a similar outlier again. And so I think you'll see deep learning systems that when they detect novelty. So I almost imagine this deep learning system. Sérgio, you think this is creepy because I want to design a deep learning model where I have like eyeballs and pupils, and so I'm feeding Data through it. And then when I feed it something novel, you'll see the pupils dilate like, Whoa, whoa. And you know, behind the scenes, this deep learning model is doing some very aggressive augmentation on that recent observation. It's trying to be as aggressive as aggressive as it can to comprehend what is this new thing that I just experienced where I think a lot of deep learning models today are quite dumb, like you give it an outlier and it's not smarter for it. And so I think the deep learning systems in the future will be very, [00:56:00] very good at quickly adapting to an outlier, even a single observation. So which is what the human brain does.

Harpreet: [00:56:05] If we think about, you know, eventually gaining some type of cognition, right? I think language models will be instrumental to making that happen because I mean, if we if you follow Noam Chomsky AIs idea,

Harpreet: [00:56:19] You know his his theory of language

Harpreet: [00:56:21] Language is like the fundamental aspect of cognition. That is how we think. So developing really good language models, right? It would maybe help facilitate

Harpreet: [00:56:33] Cognition within within, I

Harpreet: [00:56:35] Would have thoughts on that.

Speaker7: [00:56:37] I I worry sometimes that we're trying too hard to build up into the right, not knowing where we're going, and I really like the idea of working backwards like I feel like for everything in my life, working backwards is so useful. Like we've talked about value, we talked about

Harpreet: [00:56:49] The more you're in the head of the the

Speaker7: [00:56:52] Person at the end, that matters. You know, people want to get promoted. They want to be successful, they want to get their bonuses. The more you're in their head, the more you can work backwards to success. But for this example? Um, you should build an A.I. system that can learn a language, so I think you're exactly right, like a language is kind of the core of it. But when I think of language, I don't think of like GPT three, I think of an A.I. system

Harpreet: [00:57:13] In my kitchen. I just had a new baby. Just kidding. I didn't. But if I did,

Speaker7: [00:57:17] I have an A.I. system in my kitchen that is there and I'm interacting with it just like I would a child, and it begins to mimic and learn a language. And everyone on this call would say, well, just because it can mimic 20 words or 50 words like your child, it's still not there. And I think what you would see is you'd have these milestones like 20 words through experience, two year old level three experience. And even like a seven year old through experience, AI experts would still say this thing is not conscious. But when it goes to school and it's studying physics now, we start to feel uncomfortable.

Harpreet: [00:57:50] Yeah. Joe's talking about Ben, just talking about babies, just talking about one child, learning just a little. I guess that anecdote here at one shot, learning in

Harpreet: [00:57:59] Babies earlier [00:58:00] this week.

Harpreet: [00:58:01] I dark outside in the morning time to have my son with me looked out the window, point to the moon and said, Moon. And he's like Moon. And now every morning since then, he's been running outside straight up to the window and saying, Moon, Moon, Moon looking for the moon or when we're in the, you know, backyard at nighttime, whatever he's like looking around for the Moon. So one shot like this is something I don't know too much about. Is this something that like, there's actual? Research into is this something that's just like?

Speaker8: [00:58:33] I mean, the notion is it's smart in computer vision, but you

Harpreet: [00:58:36] Could use it to not use

Speaker8: [00:58:37] As many training examples and

Harpreet: [00:58:39] Still come up with basically the same thing.

Speaker8: [00:58:42] You train out a bunch of images. So I think the notion is to reduce the amount of training instances that you need.

Harpreet: [00:58:49] So is this interesting thing, is there some element of like

Harpreet: [00:58:53] Self-supervised learning in this?

Speaker8: [00:58:57] Oh, no, need to go back and read the paper, I remember reading that last year and yeah, but it seems like I mean, I think it's kind of a venomous hitting on too. I mean, this is I think this is because the amount of like GPT three consumes a ton of data and parameters, right? I mean, it's ungodly how much that uses

Harpreet: [00:59:14] Megabytes and megabytes like the thing is, I mean, gigabytes megabytes, it's

Harpreet: [00:59:19] Many multiple tens of gigabytes.

Speaker8: [00:59:21] Large, huge. Yeah. But you know, I mean,

Harpreet: [00:59:25] This is

Speaker8: [00:59:25] A field where I mean, I'm obviously more data engine, but I do still kind of keep abreast of this stuff because I think, you know, it seems like a lot of the heavyweights in the field, too. You know, you Hinton's and and so forth are trying to think of new ways. You know, I think they've they've they feel like deep

Harpreet: [00:59:39] Learning sort of reached a

Speaker8: [00:59:42] Kind of its limits, and I think they'd be the ones to know since they're the ones who kind of invented it in the first place. So but it's it's hard to say where it goes. I don't know. I mean, I would say if you wanted to know, I mean, those would be as good a place as

Harpreet: [00:59:55] Any to start to look at

Speaker8: [00:59:56] The stuff they're putting out in publications and papers because I'm [01:00:00] always fascinated by that and causal learning. That's one everyone talks about. But you know, I'd be interested to see where that goes.

Harpreet: [01:00:07] So like that, Bayesian belief networks type of thing. Ben, you're about to say something I'd love to hear from you. If anybody else has anything to say or ask on this topic, please do know your thoughts. Input questions are

Harpreet: [01:00:21] Welcome. Ben, go for it.

Speaker7: [01:00:23] I think one of the huge things that's listening or missing

Harpreet: [01:00:25] From these deep learning frameworks

Speaker7: [01:00:27] Is the ability for recall. I don't mean recall by the data science definition. I mean, recall is in memory. So search is something that is, you know, Google has search, search engines have search. But when you're building a deep learning model, you don't

Harpreet: [01:00:41] Really think about its search

Speaker7: [01:00:42] Capability.

Harpreet: [01:00:43] You can. You can you can carve

Speaker7: [01:00:45] Off the bottom of your classifier or right above it, and you can pull out a late in space and you can do some clustering with that really cool stuff. You can do search with it, too, like image, a pre trained image that vector you can search with it.

Harpreet: [01:00:57] But I think

Speaker7: [01:00:58] That's something when I think of the models in the future like, and I apologize for bringing this example up again. It's obviously top of mind for me. So if I think about a model just experiencing life through a camera, if I ask that model later, Hey, check out this frame or tip, check out this object. It will be much better than a human going back in time and in its experience in placing it.

Harpreet: [01:01:20] Yeah. So here's something interesting. So let you finish this. Sorry, sorry. Apologies.

Speaker7: [01:01:26] I think the last I thought I

Harpreet: [01:01:27] Had is just, you're

Speaker7: [01:01:28] Going to see a reduction in the number of times the AI has to review. Actually, I might be I might be going past. I'm about six. I'm going to say the number of images that it has to review. I'm saying it's going to go down, but I don't think that's actually true because when it comes to novelty, I think it's going to go up. So just like humans have to dream and we have to like form these memories and connections, I think I will have to do some of that too. But this experiential learning. I think that's what's missing experiential learning where it's like this temporal flow.

Harpreet: [01:01:58] Well, that's a that's an interesting thing [01:02:00] because so much of our I mean, we our experiences. It's sequential, it's linear, it's, you know, one through line for our experience, so let's say there's some artificial intelligence, machine intelligence, the way they process things is not going to be necessarily sequential. The way we do because they're distributed, they're parallel, asynchronous, even. That's not how human brains work.

Harpreet: [01:02:26] So I wonder what?

Harpreet: [01:02:28] You know, was an entity like that today, tomorrow, yesterday.

Harpreet: [01:02:33] It's counting, no age time means nothing to it.

Speaker7: [01:02:38] Yet it doesn't know time, but I think you're going to you're going to see some really fun behaviors of these systems that feel like they're living.

Harpreet: [01:02:44] They're going to show

Speaker7: [01:02:44] Surprise and novelty and will all chuckle on this call because we know they're not living. But if I show them to my parents, like if my parents come over to my home, my home is going to show huge interest in them the first time they're very novel, it's very interested in them. It's paying attention to the cameras, actually panning and tracking them while they walk through the house. It doesn't care about the kids or the dog or anything else. But then when the parents start coming more? Yeah, not that interested. But then when an intruder comes or like the house is on fire, some truly novel experience II is wide awake. That's why I like the pupil dilation. So hopefully in the next six months, you'll see me show up some demo of some experiential based model that has pupils dilating when I throw new things into it. It's really interested in the novelty. I did want to call it something that we talk about language. There is the language

Harpreet: [01:03:34] Of the world when you have like

Speaker7: [01:03:35] Comparisons like size, comparisons and relationships. I think this is something that deep learning could do better at. So if I turn something upside down in the frame or if I shift something, the model that should be a novel experience the novel and I know people talk about this like, imagine, you know, being able to imagine what's missing. So I think we understand the directions things will go. I just don't know if people are [01:04:00] thinking about the right way because we're distracted by these massive models. Can we can we go bigger and bigger and bigger and bigger? But I feel like they're building these like air balloons in the wrong direction. They're not kind of going back to the basics of. You know, how does my child learn anything?

Harpreet: [01:04:16] You know, let's let's hear from you.

Speaker7: [01:04:23] Your mutagen?

Harpreet: [01:04:24] Yeah, sorry. Yeah, there we go.

Speaker5: [01:04:26] Better so I don't want to take us in a way off course, but you know, I've been seeing a lot of what I imagine

Harpreet: [01:04:37] This hype about quantum

Speaker5: [01:04:38] Computing mean is the fundamental issue then, do you think? The algorithms, and if we just get more faster compute

Harpreet: [01:04:50] Power or just it's just the

Speaker5: [01:04:54] Algorithms, or do we need a fundamentally different hardware architecture? What do you think?

Speaker7: [01:05:01] I don't know if quantum computing will. I think the most profound thing I heard about quantum computing is someone who's at Google now. They said it won't make our current models run faster. It will solve new problems we haven't thought of

Harpreet: [01:05:15] Because I

Speaker7: [01:05:15] Haven't done any programing to kind of understand the restrictions. Some of the ways we approach problems might not work very efficiently on quantum computing, but I do think we're going to need an order of magnitude increase in like

Harpreet: [01:05:29] I'm a huge fan of Nvidia. Nvidia up into

Speaker7: [01:05:31] The right, up into the right like keeps running faster. But I think as we get into these much smarter systems, like even when I think about like how how would I design a model to react to a novel event it I quickly begin requiring some very high compute. So I think we are going to need orders of magnitude more compute to have an AI system that is completely useful. Like, I like to think of the smart home, the smart home

Harpreet: [01:05:57] That you don't want to fire, the smart

Speaker7: [01:05:58] Home, that you don't

Harpreet: [01:05:59] Want to delete or [01:06:00] turn off notifications.

Speaker7: [01:06:01] And I think we in the next 10 years, we will all have that,

Harpreet: [01:06:05] Which will be super useful. You can actually have

Speaker7: [01:06:07] A conversation with your home and all of your appliances will be in total command and control. And if something's annoying you, your home will know that like if you're annoyed because the house is messy.

Harpreet: [01:06:20] And the home will quickly figure out where the mess came

Speaker7: [01:06:22] From, it wasn't from you, is from a kid. And then the home will realize, well, do you want a consequence applied?

Harpreet: [01:06:28] You know, I know the kids watch TV.

Speaker7: [01:06:30] I can disable the TV. And so you can imagine this type of

Harpreet: [01:06:33] Interaction where you're like, Well, actually, yeah. You know, you see me

Speaker7: [01:06:37] Yelling at the kids, no TV because the House is messy, the home just says

Harpreet: [01:06:40] Done taking care of.

Speaker7: [01:06:42] But this wasn't like a Data science heroic effort. This was just the natural conversation. The home is not living, not alive. Was it useful to you? Hell yeah, it was useful. You're like, Take that, kids.

Harpreet: [01:06:55] Like the episode of Black Mirror, when a lady like has like a miniature version of herself digitized and becomes a smart home for her, it's a crazy episode. But even up upstream from from all the computation and all the algorithms, we don't understand how this actually

Harpreet: [01:07:13] Works, how this

Harpreet: [01:07:14] Experience actually works. I mean, we can't even model a Paramecium or Heroes. mean much less so that the human brain

Harpreet: [01:07:24] Or single neuron,

Harpreet: [01:07:26] I mean, we can model neurons with deep learning kind of does,

Harpreet: [01:07:30] But it's a

Harpreet: [01:07:32] Tough problem. Antonio says he'd go to Ben Taylor University Campus Location, Utah. Oh.

Speaker7: [01:07:40] Nice, thanks. Like my my brain spins out of control, thinking about in a good way, thinking about the smart home because like, how many of us have left our garage open? We're like, Darn it, dang it, I don't want that open in your home would very naturally. Like, I would actually like a smart home that just starts closing it for me, like it figures out [01:08:00] rules, and if it closes it too soon and it notices, I don't like that it adapts like a controller. But you can imagine a smart home that just

Harpreet: [01:08:06] It escalates

Speaker7: [01:08:07] Into like a queue, a potential automation tasks.

Harpreet: [01:08:11] Would you like me to close your garage

Speaker7: [01:08:13] When it's been open longer than 10 minutes? I didn't ask for it. It's just been learning, and it thinks that's useful. Or like, would you like me to? You know, I notice you have a fitness routine, would you like me to hold you accountable to that?

Harpreet: [01:08:30] And in what I'm saying is

Speaker7: [01:08:33] I'm not forcing you guys to be held accountable to that, but for me, I might have my own life where I'm like, Yeah, hold me accountable to that. And then, yeah, so those types of interactions, I think. They're limitless, like we could come up with 50 ideas where we would all agree. Yeah, that makes your life better, but those ideas are specific to you. And that's something I get super excited about.

Harpreet: [01:08:56] Super for use case like that. Like just exactly what you just described. Where are the limitations falling short? I feel like that's something that could

Harpreet: [01:09:04] Mean deep learning to probably solve that, like for sure.

Speaker7: [01:09:07] I think the thing that's falling short is if I if I wanted to be critical, I won't call it experimental, I'll call it brittle A.I. That's where people approach problems with horse blinders on, and they're so focused on like elderly slip and fall. I see you fall. A better way to think about it is anomaly detection or like all the data that matters. So all these systems, they shouldn't just be consuming.

Harpreet: [01:09:28] Video What time of day is it? What are all the

Speaker7: [01:09:31] Features going on like? How many people are in the frame like they will have so much

Harpreet: [01:09:34] Data, but this doesn't

Speaker7: [01:09:35] Mean it has to stream up to Amazon, where you feel violated with a privacy concern. Like even today, like I've got this A.I. system in. My seven year old was like

Harpreet: [01:09:45] Naked in the living room and like that's

Speaker7: [01:09:46] Captured in the TV. I'm like, and my record is like, What the hell like? Why are you naked? Like, why are you naked down here? He's just like on his tablet. But like, I don't want that information

Harpreet: [01:09:55] Leaving my home, but that information

Speaker7: [01:09:57] Can be processed by any system

Harpreet: [01:09:59] Like like [01:10:00]

Speaker7: [01:10:01] You guys might think this is dystopian or like, this is a feature you don't want. But imagine the AI house

Harpreet: [01:10:06] Saying get your

Speaker7: [01:10:07] Pants back on like it's the main floor. It's my kid. He's butt naked. Get your pants back on. Your father's come and get your pants

Harpreet: [01:10:13] Back on in the kid is like,

Speaker7: [01:10:15] No, and then the system turns off his

Harpreet: [01:10:17] Tablet. He's like,

Speaker7: [01:10:19] I hate you, I'll go get my pants on. Um, I don't know. I might be too obsessed about this stuff. You guys might start to wonder, like, does this impact childhood development if? If the if the rule enforcer is not a human and I'm sure it does, I'm sure there's some impact there, but whether it's good or bad is unknown, I think.

Harpreet: [01:10:41] Mean, that's something for you budding. Psychologist, behavior psychologist out there to test that out, I think that'd be interesting to test out, too bad for those kids that fucks them up, but.

Speaker7: [01:10:54] Well, I just want the laser system that kills fruit flies. Like if it ever sees a fruit fly in my kitchen,

Harpreet: [01:10:59] It'll be like game over.

Speaker7: [01:11:01] No humans are in, but you'd have to have super high trust because no one wants to go blind.

Harpreet: [01:11:07] Walking into their kitchen.

Speaker7: [01:11:10] But even just like imagine any system that prioritized what order you ate your food, you know, the bananas are going to go bad in three days like they AIs system already knows that. And it's telling you what you're going to eat for lunch. And it's like, Well, you like this sort of. But if you don't eat this, you're going to waste a lot of food. So it's like it's trying to optimize. It's like it reminds me like the traveling salesman problem is trying to like, optimize like this mixed bag of suboptimal lunches or like breakfast. And and if you haven't gone shopping to the store, then your lunch options are going to get really shitty where it's combining like. The worst possible combinations, and you're like, I can't eat this and your home is saying, I know,

Harpreet: [01:11:47] But you need to go to the

Speaker7: [01:11:48] Store. I already ordered

Harpreet: [01:11:49] The pickup because I knew

Speaker7: [01:11:51] You'd be angry and it's available today at 3:00. I looked at your calendar and you're available.

Harpreet: [01:11:57] You know, go for it.

Speaker5: [01:12:00] So, [01:12:00] OK, so I'm off mute now, I put my hand down, there we go. Yeah, so. Not to interrupt

Harpreet: [01:12:10] The kind of meta or

Speaker5: [01:12:13] Larger

Harpreet: [01:12:14] Cognitive

Speaker5: [01:12:15] Questions, computing questions, model questions, algorithm questions, more practical matter because you mentioned

Harpreet: [01:12:23] Amazon and

Speaker5: [01:12:25] Stuff going up to the cloud, et cetera, et cetera. So while you guys were talking earlier, I was on the one hand thinking quantum computing. And the other hand, I'm thinking iOttie Small

Harpreet: [01:12:38] Devices models on a chip.

Speaker5: [01:12:41] And I've heard of, you

Harpreet: [01:12:43] Know, this being a thing,

Speaker5: [01:12:46] Especially with sensors and actually also, I mean, probably in a range of areas, but for example, an

Harpreet: [01:12:52] Environmental monitoring,

Speaker5: [01:12:55] If you have sensors deployed, you know, way out in the field someplace,

Harpreet: [01:13:00] Obviously you're not

Speaker5: [01:13:02] Going to have a workstation or something and or any number of other sensors out in the real world, you're going to

Harpreet: [01:13:12] Need a really

Speaker5: [01:13:13] Fast response from that sensor to do something. So does that. I don't know if that factors into your thinking at all. Ben, I mean, you're talking about something that's a lot bigger. I'm not saying that

Harpreet: [01:13:30] Some small

Speaker5: [01:13:31] Device is going to do all these things, but maybe there's some kind of integration there with the hardware and the algorithms and so on and so forth.

Harpreet: [01:13:42] So I think you're exactly right.

Speaker7: [01:13:44] The way I think about it is all of this inference actually happens on the edge. So you have edge sensors because it's actually too much Data. So if you have 50 cameras in your home that are streaming 8K resolution, you actually don't want that

Harpreet: [01:13:56] To go anywhere really like because that's just a ton [01:14:00] of data.

Speaker7: [01:14:00] So you want edge devices that are deciding if things are actionable and then like routers, they are escalating them to the main hub. And the main hub is what is sentient or like. Some people think it's sent to you and it's making decisions. There is an evolution to like you have in video in companies like that that are trying to make certain A.I. models run faster, they're trying to change some chip architecture. But the ultimate is to actually make custom silica where you're you're saying, Hey, for this particular use case, we will actually engineer this entire chip. It's actually incredibly limited, like it can only handle this type of video input. It can only do this task. You have no option to program it. And so I think when you get there, then we will be very happy with some of the power consumption because like because everything I'm describing, if I have to run that with all these edge GPUs, even with Jetsons, you're like, OK, like that's actually showing up on your power bill now. Like doing all this massive inference in your house is a little hotter because of it all.

Harpreet: [01:14:56] So I think you're exactly right.

Speaker7: [01:14:58] I think we will get there where you have very specific purpose built. One of the things that surprised me is how easy it is

Harpreet: [01:15:05] To roll out a new chip because you have like

Speaker7: [01:15:06] These global foundries. So someone on this call, like if you had funding you had, I'm not sure what

Harpreet: [01:15:12] The number is. The numbers keep coming down like,

Speaker7: [01:15:14] Is it 50 million or 10 million to develop? You don't have to have a fab. You can actually just say, this is what I do. This is the chip architecture. You pay some consultants, they go, send it out to TSMC in Taiwan, and here's your chipset

Harpreet: [01:15:29] That you can start getting on.

Speaker7: [01:15:31] Like look at companies like Tesla, like they don't have a fab, but they can get these chips made. That's something that's always blown my mind, like what humans on Earth designed chips like, I would look at these NAND flash chips and they look like cities. You zoom in on like a scanning electron microscope and they it's like New York, but probably more, much more complicated. And there's some human somewhere that

Harpreet: [01:15:56] Understands all of

Speaker7: [01:15:57] You know, they can

Harpreet: [01:15:58] Explain all of it, which I always [01:16:00] thought that was

Speaker7: [01:16:01] Insanity. Deep learning is so easy compared to that.

Harpreet: [01:16:05] My brother in law works at Apple, he designs the chips there, so he does like chip

Harpreet: [01:16:09] Design and chip speed testing

Harpreet: [01:16:13] Insane. I don't know how

Harpreet: [01:16:14] He does it. He studied electrical

Harpreet: [01:16:16] Engineering. That is nutty. Yeah, yeah. Speaking of silicon, the M1 is insane.

Harpreet: [01:16:24] I was running a

Harpreet: [01:16:27] 20 iterations of Bayesian search on my M1 on four million road Data set using a variation of like GVM.

Speaker7: [01:16:37] Speaking of silica, here's some NAND flash.

Speaker5: [01:16:42] Nice. Wow. Got a wafer there?

Speaker7: [01:16:44] Yeah, that I cracked it. This is my good job, Ben. We're proud of you, wafer.

Speaker5: [01:16:51] That's like 12. Is that a 12 inch wafer?

Speaker7: [01:16:55] Mm. I think is that 12 inches? I think it's yeah, I think it's 12 inches.

Speaker5: [01:17:00] I don't know if they're all 12 inches nowadays or if for smaller applications they're doing, you know, like they're still using the equipment to make smaller wafers. It's it's been a while since I was knowledgeable in this area, but

Speaker7: [01:17:17] They're all about

Harpreet: [01:17:18] Scale. So I think

Speaker7: [01:17:19] Three hundred and fifty millimeter was the norm when I was there. I think they're pushing for four hundred or fifty. Like the bigger they can make the wafers, the more. It's so funny because people get so excited

Harpreet: [01:17:30] About like, Oh, the chips are

Speaker7: [01:17:31] Getting smaller. Everyone wants to make them smaller, but the economics are they cost less chemicals to make and so smaller is a bonus. So I thought that was funny because I thought, I'm so happy they make a a terabyte micro SD. But the reality is like, yeah, it just took the economics make sense. So they will always go smaller because it's it's less chemicals in material to make it. But now we're now there at like the electron limit where it's. [01:18:00] There's, you know, everything super problematic now because they're less than 10 nanometers and. You count the electrons in your memory cell.

Harpreet: [01:18:10] We're just getting started with all this stuff, we are at the beginning of infinity

Harpreet: [01:18:14] With

Harpreet: [01:18:14] Everything.

Harpreet: [01:18:15] We are really at

Harpreet: [01:18:16] The beginning of infinity. My friends. Let's go ahead and wrap it up, man. Thank you guys. So much for hanging out, Ben. Thanks for bringing on the interesting discussion. Really enjoyed that. You should listen to the interview.

Harpreet: [01:18:25] I did everybody for tuning in.

Harpreet: [01:18:28] It's not really on the podcast yet, but it's live on YouTube. The interview I did with Marcus de Soto.

Harpreet: [01:18:33] He's an

Harpreet: [01:18:34] Oxford professor. He wrote the creativity code. I love that conversation. Like we talked a lot of philosophy of mathematics and things like

Harpreet: [01:18:42] That, so definitely tune in to that. Spoke to the

Harpreet: [01:18:45] People's Data scientist himself, Danny Ma, yesterday. That is on YouTube as well, but it'll be dropped later on the podcast. Shout out to everybody that joined in on LinkedIn Antonio Greg. Good to see you guys there. Shout out to everybody that came in today.

Harpreet: [01:18:59] Be sure to

Harpreet: [01:19:00] Tune into the interview

Harpreet: [01:19:01] I released today with Andy Hunt,

Harpreet: [01:19:03] Author of

Harpreet: [01:19:04] The pragmatic programmer.

Harpreet: [01:19:07] Legendary author of Pragmatic

Harpreet: [01:19:08] Programmer has a great

Harpreet: [01:19:10] Conversation as well.

Harpreet: [01:19:12] What else I got got a bunch of

Harpreet: [01:19:14] Really interesting episodes coming up. If you guys don't mind, let me just tell you what's happening

Harpreet: [01:19:17] In the next few weeks, in the next

Harpreet: [01:19:19] Few weeks, who I got, I got interviews coming up

Harpreet: [01:19:22] With.

Harpreet: [01:19:24] Let's see here I've got. Next week, I'm talking

Harpreet: [01:19:28] To Christina

Harpreet: [01:19:31] Jacomo she is an industrial philosopher that's going to be great. November 12th that the episode with that. George Farah Khan November 19th. Steve Carindale about turning ideas into gold. November 26 Talking to Karrueche

Harpreet: [01:19:45] Alexeyeva, we talk

Harpreet: [01:19:47] About NLP and philosophy. Then December 3rd Christian Espinosa. We talk about why you shouldn't be the smartest person in the room. Then, on December 10th, talking to Dana [01:20:00] McKenzie, who is coauthor of the Book of Why. So we had a good conversation there, so I hope you guys get a chance to tune in. And if you know, if you haven't had enough for me this month, realize that there are about one hundred and ninety three episodes published that's over, you know, probably three or four hundred hours of content at this point. We're just 5000 downloads away from one hundred thousand downloads. So if you haven't listened to any podcasts from the @TheArtistsOfDataScience, go make up for that. Help us crush that that 100k

Harpreet: [01:20:36] Took, you know, a little

Harpreet: [01:20:37] Over a year and a half to get this far. Hopefully, we get there by

Harpreet: [01:20:41] The end of the year. I would love to see that happen.

Harpreet: [01:20:43] You guys take care, have a good rest of the evening. Happy Halloween to everyone. Thanks for coming by. Remember my friends who got one life on this planet? Why not try to do some big cheers, everyone?