Hi, welcome to the Artificial Intelligence Machine Learning and Data Science Weekly Podcast. My name is Kwan Hum or you can call me K-H. In this show, I'll be talking to AI-ML and Data Science Practitioner around the region. In each episode, I'll dive into relevant and interesting AI-ML topics where you get to know more about topics ranging from AI-ML adoption, best practices, and tips and tricks to be a better AI-ML Data Science Practitioner.
Hi, everyone. Welcome to another episode of AI-ML & Data Talks Podcast. In today's episode, I'm super excited to have Jason Mayes, who's currently working as a web-ML leads at Google as a cast for the show. Hi, Jason. Welcome to the podcast. Thanks for having me excited to be here today. Since you are on the tour or the death first, I think it's my managed to catch you to become a guest of for my show. Welcome to the show again.
Has the flow of my show usually asked the guests to actually do a self-induction about themselves? Maybe you can start to give us a brief background about your childhood or maybe your education, your current career path, at least you to become a web-ML as a list. It's quite a journey, so let's get started. I grew up in England and I grew up in a very simple family back there. Before I even knew about technology, I wanted to be in the Air Force back in the
old days. Unfortunately, I didn't quite work out due to medical reasons and things like this. My uncle put me my first laptop when I was around 12 or 14, something like this. I started learning how to program in Visual Basic. This gave me superpowers back then because I could admit the computer, do what I wanted it to do and get it to respond in certain ways or call some API. That triggered my inspiration to go deeper into technology essentially. I took a
gap here to earn some money and then I sent myself to university after that. I went to university at Bristol and studied computer science. I realized that I couldn't program well and I probably had to program at this point to learn the basics of languages like C, Java, JavaScript and a few others as well. I also realized that evenly make the best backend service
without a good frontend and UX user experience. People won't use it. I became very passionate about web technologies because it allowed me to do the backend algorithms and databases and all that kind of fun stuff. I also mixed it with the user experience and user interaction and all that kind of stuff for design aspects. Web engineering became my passion, if you will. I actually joined a startup called XMOS, which was a semiconductor company. I lived there web engineering for three
years. I wore many hats. I had never come in trying out startups early in your career. I get to wear many different hats. I did coding, design, UX, even photography and 3D modeling, all this kind of stuff. Eventually I realized it was time to start my own company. Alongside that job, so 9 to 5, I did the XMOS job and then 5 to 11, I did my own startup company with a web agency of sorts with myself and a co-founder. Over the course of a year, I bought up a decent
reputation about getting work coming in and it was taking off quite nicely. But also at point, Google found me because of this success and asking me to join them as a web engineer at Google London. So I put my startup on hold and I joined Google 12 and a half years ago now. I was a web engineer and I learned how to do all the scalable systems on the Google Cloud and I learned about security
and all that kind of fun stuff too. I realized the thing that really got me up in the morning on Monday mornings was starting with a blank canvas and solving a problem that hadn't been solved before. So at Google, we have this thing called 20% projects where you can take one day, week to follow your passions and learn something new or things along those lines and it was in these 20 percent times I was making these fun projects that were world first in the web industry and I
became known for that internally. So I then joined the team over in California called the zoo and they were like creative think tank inside of Google to make world first for our largest customers globally. So I got to use all those skills I learned before to apply to that kind of realm and that's where I got into machine learning and more emerging technologies because we had to like work of those
emerging technologies to help the story of our customers in some other way. So maybe there's a new Hollywood movie for Star Wars or something I might propose having a life size stormtrooper with projection mapping which is hooked up into our speech recognition APIs and natural language processing to then allow people to talk to it at a red carpet event and get answers
about the Star Wars universe or something like that. So we come up with his ideas and that's when I realized I got into machine learning because I was trying to solve a problem where I was
trying to recognize Red Bullcans and regular computer vision just wasn't cutting it. So back in 2015 just before TensorFlow has announced the public, I was able to try internally and that's when I had my first taste of proper machine learning to do image classification and I found that it was really hard to solve this even with all the knowledge I had around me of the people who made TensorFlow it is still really hard to use at that time. So I took it up on myself to take this knowledge I had
learned from my Google peers and make a system that was easy to use. I called this ZOOML back in a day which is a ZOOM machine learning technology and what I allowed it to do is you upload a video of some object that you're interested in and I had to process that with the TensorFlow library and produce a model that could have been recognized by our JavaScript API. So you send me an
image I'll tell you if it's in the image or not. That went on to become Cloud AutoML Vision and that's when I realized I want to do more of that kind of stuff, make stuff easy for everyone to use and it so happened in 2020 the TensorFlow JS team was looking for someone with a background in web engineering which I had and a passion for machine learning which I was starting to get. So I was able to combine my two sides of the coin there to become developer relations
in GDF4 TensorFlow JS and that's where I am today. Wow, interesting journey. So I think like you mentioned I'm sure a lot of people have heard of TensorFlow which is one of the most popular deep learning framework out there. But then you were seeing that there's something called TensorFlow.js Yes, TensorFlow JS is different from TensorFlow actually. Great question. So as many of you have heard of as Python TensorFlow out there that came out there first in 2016 or 2015.
Shortly afterwards we created TensorFlow JS which is the same library but written in JavaScript. So that JavaScript users can actually benefit from this ecosystem too because traditionally machine learning has been a very Python focused area because of academia that 70% of the world use JavaScript.
So we realized that as a missing gap there so some researchers at Google created this to at first put models onto the browser to put people to try and understand how machine learning works but it turned out people found this really useful and it actually then graduated to a fully fledged product on its own. So because we also know there's something what we call on device machine learning. Yes, I think there's a lot of people are very excited to actually deploy the
project at the age computing. So that's a lot of people to do that also because if it's just can run on web browser the education definitely run on the edge. Yes, so TensorFlow JS is an interesting one. We've got two incarnations of it. There's been no JS version which is just like Python on the server side and then what the JavaScript version that runs locally on the client side. So by running locally that gives you the privacy, it gives you low latency and of course lower cost
because there's no cloud GPUs for the inference being used. So those are superpowers by running on the client side. Yes. So I'm sure a lot of people have seen Jason promoting TensorFlow JS project over the internet and you have a lot of interview. You have interview with many people, many developers actually showcased their TensorFlow project, TensorFlow.js project. Yes. Any project that you that actually blow your mind when you see it that
I still remember I do today. I think all the projects have blown my mind over the years as they keep coming out better and better things. But one application that I particularly like is the application of TensorFlow JS and the healthcare industry. There's one company called Include Health that does remote physiotherapy using our pose estimation models. So if nothing more than your regular webcam in the browser you're able to understand the 3D pose of the human to understand for physiotherapy,
the range of motion and all this kind of stuff. So now people who might find it hard to get to a doctor especially when you've got physical issues with your body can actually do it from the comfort of their house in a privacy preserving way and the aggregate results of the range of motion are sent to the doctor rather than the raw imagery necessarily. So their privacy is preserved as well. So this is a very interesting use case and I think we'll see more of those happening
into the future too. Yeah. Looking at the pose estimation model there's a lot of people actually use the model to actually do a lot of stuff for example how to play golf, how to have a proper yoga pose. And recently there's a lot of project that actually incorporate other libraries for example the MediaPap libraries. Yes. Maybe you can you elaborate more on that? Sure. Yes. So alongside TensorFlow JS there's also a team at Google MediaPip and they're most famous you know
for their body based models. So things like face key point detection or pose estimation or body segmentation, those are the most popular models probably. So very about understanding the human body and some shape of form for multimedia applications. And their team has worked closely with a TensorFlow JS team in the past to port some of their models to run in the browser as well. So many of those models can now run in the JavaScript environment by technologies like Web GPU or
WebAssembly to get really fast performance in the browser. Okay. So where do you see the the next trend on the next upcoming things that can be done with TensorFlow JS and MediaPip? So we're starting to see in 2023 the Genai stuff being pushed to the client side as well. About a year and a half ago I saw the first version of a diffusion model being run entirely in the web browser. Of course you do need a large graphics card to do that but if you have one
on your machine then you can run it locally which is really cool. And we're starting to see the same pattern happen with the large language models as well. Obviously the very large language models you're going to need a very high spec machine to do that. But I think we're starting to see a trend
of shrinking these models down. So everyone got excited initially that how large can we make these things but now people are trying to figure out what's the minimum I need for a certain business application and of course in the real world when you're trying to make a profit and make money you want to be efficient and not waste a lot of money on the cloud services to run those LLM.
So I think we'll see a trend of these being made or distilled if you will for certain subject areas maybe like legal document summarization or code generation or things like this and these individual models will be downloaded as they're needed and cash locally to then perform a task for that user on that website. Yeah. Yeah I think there was a time that people all rushed to actually build a larger model, be more parameters and train the larger model but soon they realized
that actually now they try to condense the model. They find that actually if they have a if I have a condense model with the lesser parameter it actually performs reasonably well also. I think that's the thing that people are trying to do. How to see how to actually push a model that actually lesser parameter but perform reasonably well also. Yeah and I think that's like a wise thing
to do generally with machine learning. I see a lot of people ask me where they're saying should I use machine learning or regular computer science and say well if machine learning actually gives you a significant edge over the regular computer vision techniques then use it because explain
abilities obviously better as well. So and the same applies to these LLMs and things having a smaller more concise model is probably better for for many reasons then to just have the biggest thing that's maybe good at everything but you're only leveraging a small part of it really in
day-to-day usage. Yeah. Interesting. So talk about our I'm sure that a lot of people are very curious about how does LLM is going to affect the work as a data scientist or machine learning engineer or AI scientists because people think that now we have all these gikub co-pilot or do it AI or whatever LLM as AI assistant model out there that can actually help us to do programming and they were saying that sooner or later they might not need to be a job as a programmer. Sure.
Do you feel this thread? So I would disagree with that in the near term feature until we have like AGI or something like that. I think right now it's good as a co-pilot so if you're stuck on something it can give you a suggestion that may or may not be right but you as a human can look at that and maybe get some inspiration from it. So using LLMs or even diffusion models for inspiration in your work I think is for sweet spot right now completely relying on it I personally would not do that.
If you look at some of the code it produces it might be functioning correct but it might be not very efficient or not very easy to understand or read so there's other aspects of software engineering but it doesn't quite get right now and even if it does it could hallucinate answers as well which might not be correct technically speaking so I think as a means for inspiration is the best place for it at this point in time. Yeah. Okay so to all the people who are worried it's not that bad yet.
So I think another thing they always boggle my mind is so we know that LLM model is trained or LLM of data that's available out there but now with LLM model it also generates a lot of data. So do you see there is a visual cycle where generated data which is not correct like you say hallucination by all these LLM is going to be no feedback into the model again and going
through all these process again. There's definitely a possibility and I'd also argue that a lot of humans incorrect most of the time as well so I think when they're better off when we were before in that sense we just got more of this stuff being generated and at a faster rate and I think this is a journal trend in human history that we're producing more data than ever before and it's going to be more important for companies like Google we're about to try and fact check this stuff to make
sure it is accurate as the best we can. So I think we'll see more research going on in those areas into the you know not too distant feature to try and make sure that the stuff it's coming back with actually can be verified with a trusted source but it is probably true and it's not just hallucinated in some form. Yeah. So basically in a way that the more research need to be done therefore there's a more opening or job opportunity for people to do a computer size or machine
learning works relatively to actually a good model. Yeah. Okay so the next thing is anything that you are passionate outside of your work actually we would like to find out. Outside of my work. Because I see for a profile you call yourself as a hybrid developer or design technologies what does that mean actually? Yes so I've always been a strong believer that in order to innovate you need to cross multiple industries and take inspiration
from many areas right. So I've always defined myself on my website if you look at my website currently at least it says I'm 51% technical and 49% creative so I've got a slight bias of a technical side being an engineer but I also do appreciate for design and UX aspects as well. So I've always tried to blend those two things in anything I create so I won't just make research for the sake of research I'll try and solve some problem or make something more efficient when it
was before. So it has some business value or value to human life or something like that. I see a lot of research going on these days it's just for the sake of research rather than actually getting at something. So I've always had that spin on things and I think if you're looking to stand out in the crowd as well it's good to be a hybrid of two industries so maybe you're interested in music but you've got a passion for machine learning then focus on like machine learning for music or
something like this and you'll be able to find a niche for yourself where the job you're doing isn't a job because it's just play to you right and that's really like the it might take a few years to get to that spot to get to that perfect role but if you aim for that I think that's a good a good thing to do because Ben well I don't work a day in my life right now I'm here talking to you on a podcast and this is part of my role now which is awesome so I also get to play with a lot of new
technologies and experiment and innovates. So innovations one of my side hobbies I guess I love to make things in my spare time sometimes I have coding dreams and I'll code something in my mind when I'm dreaming and then over the weekend though I just implement it and it works which is cool so
why you mean you do you know in your dreams you do coding yes it happens wow something do you have it do you do you back course in during your dream sometimes I come up with my solutions like that so there's one situation where I was trying to make this teleportation demo where I was
able to teleport myself using WebXR and body segmentation and other things WebRTC that kind of stuff and I got stuck on some problem when I was making this but after dreaming about it for a little bit I actually came up with a solution and then I actually tried out on the weekend and it actually
solved the problem so it's it's cool the subconscious can work in mysterious ways sometimes say yeah I missed your mind has no rest during your trip I'm not happy for my effect it's still doing a lot of coding work so I'm sure that you have a deal with a lot of people working in machine learning
projects and also coming to the part where the deployment of the machine learning project what do you think is the biggest challenges or that people face when they want to deploy machine learning project shot and of course it changes from person to person but the main common themes I see are
how do you scale so a research project might be you in 10 teammates but how do you scale to tens of thousands of people in a way that is cost effective and obviously some of the solutions on Google Cloud can help with this by using like TPUs or something if you are on the server side
to reduce the cost a little bit there but also security as well I see a lot of people who are early in their careers make web services and things that are very easy to hack and obviously this comes with experience you learn how to adapt and prevent these things in the future but I think
having more more services that allow you to deploy in a single click that solve these problems for you will help these people out and there's whole industries working on that right now to make machine learning models one click deployable so that you don't have to worry about scalability
and security and it just solves all that headache for you so I think there'll be an ongoing research there and it's good to know about yourself because you can also save money by not using those services and doing it yourself sometimes too so yeah so coming from a animation background so I
think that like Malaysia we just that that offering data science and machine learning causes in the university universities yeah so we talk about this problem like scaling wherever it's kind of a very practical problem but it's very hard to actually explain this problem
tournically in the universities so do you see that for you city program wherever they need to be more connected to the industry so they can actually showcase all this relevant problem that the student might need to know you've got to if I'm just starting from textbook I might not
I would see all these problems sure that's a very good point one thing I really liked about my degree course when I did computer science is that they forced us to turn our research into a business even I did not study business computer science I did pure computer science but we had to make a
business plan and we had to make a fully working product that you know theoretically could scale and all this kind of stuff so at first I said why are you making us do this but I can't thank them enough now later in my career because that actually is very valuable knowledge to have and Bristol
University at the time had some of the most highest number of startups being formed because of this extra element they had in their degree courses and I think it's important for academic institutions to realize that a lot of people will not stay in academia they'll go on to regular
real-world jobs rather than just staying in academia and there's other problems that must be solved there or at least known about in order to be successful in that environment so I agree of you yes I think they should teach these things alongside this is how you make a great machine
learning model but this is also how you deploy it at scale yes yeah interesting because a lot of time coming from a lot of a commission they are very I would say perfectionist they like to they like to come with a perfect model but actually in the real world deploying a model making
the model works and making the the r-right of the model is more important so I think if if any students or whoever coming to the industry if they have some business sense like you what you are seeing then how does the model actually benefit the company that's very important because no point
spending a lot of time where the model is still being researched and it's not deployed to then the model has not been utilized there's no benefit being actually brought into the company sure yeah and I think we see this a lot in business in the business world over the years like many
very intelligent people come up with a piece of research but don't see the business potential and it sits on the research shelf for many years and then someone like a Steve Jobs of the world sees that research and just changes it slightly into a more form that people can understand and use
maybe add it with a better UX or design and then you got yourself the next iPhone or something right so I think it's important to not miss those opportunities if you just give an extra mile and think about these other things like UX and design and the business application and how
and why it solves the problem for someone more people have that chance of being successful than as well I think so yeah interesting because you you you are very focused on the UX and UI yeah the user the different user experience do you see that being a computer science or
machine learning or AI scientist nowadays the experience itself that the the the product being deployed is so so quite important or action or characterised it's actually quite important actually yeah I think so this is one of the reasons we this year we launched a product called Visual
Blocks ML which is a local drag and drop interface for using machine learning models and connecting them to various sensors and outputs so you can go from idea to production faster and this actually wants them awards at the Kai conference for HCI and the reason I bring this up
is that a lot of research models just sit on the shelves and that's why we made this product so more of its research can get into the hands of people to try out an experiment with and then innovate so I see a lot of great research written up in research papers but very hard for the
average person to understand even though the average person might know I need a segmentation model so we know what they want but we look at the research paper they have no idea like what that actually means but we just want to be able to take it and try it out so if you think if you can
enable that for people I think we'll see a lot more innovation happen within society generally speaking and you can like take model A from person B and get the best and C's model and combine them together to do something even better than they could individually so I hope more of that happens
organically in the future yeah so I did that's one thing that like about Google has a lot of things that actually can be used and deployed can be prototype very fast for example with Google Collab you can actually try out Google Python code just using a browser or now we have this visual
blog we can actually try out building a simple prototype just by dragging it down then like now with large language model Google has maker suite where you can actually prototype a large language model yes that's right easily as much as possible now even Google has a project IDX
where you allow to actually code similar to visual code a VS code but just using the browser exactly and all these themes for common theme between them is the ease of use and the same for Google itself when we before Google search there was Yahoo search and ask jives and ultra-vista
but the thing that made Google popular was its ease of use one simple text box you put what you want and magically come back the answers right so it always comes back to UX and understanding the human in the loop there in order to be successful yeah so do you do you think that for any or
this conversion the design is they should also introduce a subject at least as students must know UI UX I mean some people are just not interested in that aspect and that's fine but they should at least know we should be aware of that this exists and if even if they they want to do it they
should maybe loop someone in who does know about it to help them polish up their final API or whatever they're exposing to the world to have the best chance of it getting used by people I guess so I think that would be good to have maybe the computer science department can collaborate with
the UX and design departments to work together for at least one project so if you understand how to work together better in the future yeah so I see this this this thing that the people are missing out a lot especially because I've been the judges of a lot of hackathons yes or even when
when I see people pitch for a startup so a lot of time people who if I'm a pure technologies I'll focus more on the model talking about model but when people look at the product and look at the the UI UX you know how the product makes me feel good yeah so if I'm not presenting a good product
that makes me feel good I know better how good is technology behind I might not want to use it yeah yeah so I think that there's something with always lacking so I see that the people who actually want hackathon or people who want pitching to get funding for any startup this UI us if components
usually are lacking actually most of time yeah and I think this also comes from my own personal experience where I'd be in both a backend engineer and a front-end engineer and I made some very complex things that solve some hairy problems on the backend and you get like a yes come out on
the command line and you're very proud of yourself but the end user doesn't care what basically is the button change on the UX and that oh well you may progress okay and that's when I realize myself that these things matter just as much as the backend say yeah okay so how do you see
let's say I would want to I'm because nowadays we have a under we are in the desert deathfest season now for the our listener our day Google has an annual event where we have a big event where we have the right speakers to actually share the latest technology
to our audience which is deathfest so Jason actually came to Malaysia to become a speaker for deathfest a Polymple and deathfest Josh now yes do you see that attend this like this benefit from this event and then how do you think that what what can take away that a student may attend
for all this event what can take away that they should go back and know think about for more this event so I just finished one one talk in Malaysia say far here in Kuala Numpur so the feedback so far has been pretty positive from the people who added me on LinkedIn and sent me
some feedback say a lot of people were very thankful that they were able to realize where they existed is quite new versus the traditional Python approach so I think people just didn't know it existed before so and that's my job to go out and tell people it exists and to go try it so
my hope is that they will go and try it afterwards not just attend the talk but play with it a little bit and see if anything that I said in the talk might be useful to help solve the problems bear dealing with in their you know maybe the personal lives or even their startups or companies
that they worked for so they can apply any of that to those situations and if just a few of those people managed to I think that's a good success and we might see some great stories come from that and maybe they'll be on my show and tell in the future on YouTube so yeah should be interesting
so basically I think all the deathfest attendees because the topic is so what there are so many topics that have been discussed in that deathfest event like for example in the deathfest column poll I think we have around 30 speakers so the topic is so what but I think the students
attend this attended the event what the main thing is that hopefully they will learn some new technology and go back be excited about it and then be curious about it and go back and try something out and then hopefully our mission is done then we are able to you know
uh give them this this this opportunity to learn something new and go back and try something else yeah I agree that's me so what coming to the wise of uh because you you coming from a background of a computer science student but then you actually went and do your own startup and
then you join Google so you have seen your started your own company you also join a Google if I would to join if I would to be interested in data science now or I'm not for in the computer science industry or from the background yeah so where do I get myself started how do I
yeah how do I how do I get myself into this industry yeah great question um so obviously depending on your skills and expertise you might start in different areas but generally speaking you're going to just start somewhere the hardest thing is to get going get some momentum going
and you know if you start at the very low level straight away you might start crying because of other mathematics and things involved in machine learning but you don't have to start at level you can start high up maybe find a model that interests you like the LLMs or um a post estimation model
and just just play with it on the high level just use it as a black box and uh understand the limitations of that model and try and um apply it to some situation you want to solve once you've done that a few times you can then start peeling back the layers to see okay how does that model actually
work what what makes up that model you might find it is actually several models contained in one model in order to make it work properly and then that sends you down a rabbit hole of learning essentially and you can then start to specialize in one area so maybe you're more on the vision side
so you specialize in computer vision models or maybe you want to do a natural language stuff the choice is yours but um go wide first see what's there pick a few models that you like the sound of and might be useful in your industry or career and then go deeper and refine your skills
in that area um but the main thing throughout all of this is to keep producing interesting pieces of work that showcase your ability to use these things um because if you don't do that no one knows how good you are at that especially if you don't have a degree in that subject so you need
to prove that you're as good as someone with a degree in machine learning these days I mean machine learning was even a degree back in my days I don't have a machine learning degree so all I got going for me is my reputation and the things I've produced over the years to show
that I can use machine learning well enough but I am credible to talk about it um so same situation takes time to build up that and one other way you can help build credibility is to use like you know LinkedIn and places like this to talk about the things you found discuss this with your peers
and share other people's work and give your commentary on it like why is it important how can you other people use this and show demos that you've created of those APIs that someone else might have made but you're starting to use those and build your reputation in that area too
I think a lot of kids these days I see forget that aspect of um business networking um in the business world at least in America it's very important to have a strong network and that can often lead to very cool career opportunities from that um about five years after
me graduating no one asked me about my degree anymore it's all about what's your experience what did you do what will you show me your work uh back kind of stuff yeah yeah I think uh so uh so I think like any anybody who actually learned new things uh keeping the momentum is very important yes
yeah uh sticking the first step is always hard and then uh trying to come up with uh uh uh uh uh uh taking the first step and then keeping know the momentum then there is a uh showcase point you know every every time uh when you when you you have achieved something showcase whatever you
have done yeah because when you showcase your product to the people like uh showcase in the kick-up or like a block above it or or or showing in thelink in whatever you actually gain some sort of like people who say it okay they actually kind of like agree that you have done something. And then also you have probably received for comments or criticism.
Yes. But it's good because that will allow you to actually improve further and then you be one from that you actually go on and then you improve your skill set along the way. Exactly. And I think what Jason said, that work is very important. Yes. So I think a lot of our audience in Malaysia, like students in Malaysia, they are very shy. They are very timid.
By joining all these like their first event or going for a community's event, if you can, try to also know, bring up yourself to become a speaker. So when you speak in the event, you actually need to actually prepare yourself. By preparing yourself and then when you go and give a talk, you actually improve your knowledge and also you improve your networking skill. People know you better. 100% agree. In fact, I think there's a famous thing called the four stages of learning.
And the first stage is knowing that you don't know anything. And you're seeking advice from experts in the field or reading a book. The second phase is you're thinking everything because you read that one book and you get overly confident. The third stage is realizing you're not as confident as you thought you were. And you go for another learning phase essentially to really refine your skills to a higher level. And then the fourth stage is mastery where you're then teaching yourself.
And everything in life goes through these four stages. Learning how to cook or learning how to program is the same four stages of development there. And it's very important to recognize when you're at stage two and not stage four because it's a dangerous place to be. And myself included, we're all human. And when I graduated from university after four years of computer science, I thought I knew everything.
And of course, my first web service, I made, got hacked and I realized, oh, no, I need to learn about security and scaling to like 10 to 1000s of users and all that kind of stuff. And this just comes with experience. It takes many years to gain that experience. It's not something you just learn in a bootcamp. It just takes time to make 20 projects and go through the vanity gritty of that and have those learnings. You only really learn from failure.
So you have to fail a few times to really realize what you're missing. And also, when you give a talk, you really better understand what you don't know because someone will ask you to explain something. And it'd be such a simple question that is essence, but you realize you can't answer it in an elegant way. And then you know you need to do some more research and you need to go a bit deeper. And it's okay in those situations say, I don't know. Don't mix stuff up in talks. Just say, I don't know.
I'll find out and give you a link later or something. But Alicia, you've been progress yourself as well. I think a lot of people these days are scared of saying they don't know. And that's also an important topic to bring up, I think, because it's only by realizing we're stuck or wrong that we can progress as well in terms of innovation or making progress on something. So it's a human trait to not say you're wrong, but it's important to realize when you might be. Actually, yeah.
Yeah, I think if you accept that there are a lot of things that you do not know, when you ask, you receive all these questions also good back. It's also good for you to go back and find out more about it. Yeah, exactly. Yeah. So, another thing that the current trend is no longer people are doing all this machine learning model training in a local premise. People are moving to the cloud.
Do you see that anybody who has bokeh in the current industry now, their at least must have a cloud knowledge or at least know how to use one cloud platform where there's GCP or AWS or Azure Yver? Many terms of machine learning, I think a lot of the services that machine learning engineers use are these very easy to use interfaces like Google Cloud, which have AI scalable systems. You just drag your data into a bucket and whip up an instance and go train on it.
So a lot of that complexity is done for you, but yes, knowing some cloud service like that is definitely an advantage. If you can go deeper and build your own, then you're even less dependent on one thing. So if something does get deprecated or something, you can always pivot very fast then too. So, but again, you don't have to learn it straight away. You can start on the high level.
And as you go deeper into making your company or business, you start to rewrite layers of that yourself to have it in-house so that you know how to fix it when things go wrong and things like that. But yeah, start high and go lower levels you need to. Yeah, I think during the beginning time, you can use all the services, but after a while, you realize that you actually incur a lot of costs. Yeah, yeah.
You need to optimize your code and then start to not deploy your services and then to actually minimize the cost and still make sure the product is still running. Exactly. And it's not easy to scale things well. But horizontal scaling, vertical scaling, there's always different ways of doing things. And on the database side, on the statics, assets, and so many different things, you need to scale properly.
So I wouldn't just, you know, if you're an ML engineer, I wouldn't just try and solve that on day one. But start to learn these things as you get more advanced and it can help you in the long term, for sure. Yeah. Okay. So just know you were the major about web AI. Yes. Okay. Maybe can you elaborate more about that? Because I know something new actually. Yeah, yes, yeah. So web AI is basically the art of running a machine learning model in the browser on the client side.
So once it's downloaded, you don't need a server to do the inference. So right now TensorFlow.js is one of the popular libraries that enables this. We've also got media pipe that have some models that run client side. And of course, the Chrome APIs themselves, like WebGPU and WebAssembly are the technologies that power all of these systems to run VAST. So the main three technologies are WebGL, WebGPU, and WebAssembly. Yeah. These are the technologies people are using to do that.
So is it something that's a library that I can just call now or do I need to learn something new? Yeah. So if TensorFlow.js library is very similar to Keras. So if you've used Keras on the Python side, we've got almost the same function definitions and everything. So you can create layers models in just a few lines of code, just like you're doing Keras, train the model, outpops the train model. You can then save that somewhere, host it, and then use it in a real web application.
So yeah, it's pretty easy to get started with. And I've got a course online that people can take if they're interested in that. Okay. Yeah. The cost is available online. Yes, it is on Google developers' YouTube channel. Google.GLE-LURN-WebML. It's probably taken a week or two to finish it, but there's 46 videos there. You can create your own time.
And I'll teach you how to use TensorFlow.js, how to use our pre-made models, how to do transfer learning, how to make your own models from scratch, from a completely blank canvas, and how to convert from Python to JavaScript as well. So with that knowledge, you can be pretty productive, pretty fast, and start making things that actually solve problems. Interesting. I'll put the link to the show description. Awesome. It's there. So is there any interesting books that you recently...
Oh, wow. Yeah, that's... I think one of my favorite books of recent past is David Eaglemann's Incognito, The Secret Life of the Brain. It's actually a book on neuroscience, but I often find myself being inspired by the neuroscience side when I talk and think about machine learning aspects. So that's a great book, and it's very easy to read, but it's got all the scientific papers, repents, and the footnotes. If you do want to go deeper and see a more nerdy neuroscience paper, as you can.
But for people who are not into neuroscience and just with a high level understanding, it covers a lot of subjects about how the brain works and what you're missing even on a day-to-day basis. If you look into a mirror and you watch your eyes move, you don't see the move. You see the resulting position. So always like fun little facts like this. But get your thinking like, okay, how is our brain actually working? What is it leaving out? What is it including?
It makes you think about how you might apply that to a machine learning algorithm or something like this, maybe. So I find that kind of stuff interesting to read. Maybe we can actually wrap up the... So if I were to be applying for a job, to be a data scientist or machine learning engineer, AI data scientist, what are the tips or tricks that you will give a wife to... So to make sure that they get the self-prepared to at least go through the interview with success.
I can't... Number one, build your portfolio. Have a strong GitHub presence and a personal website, demonstrating your capabilities and what you've made and the impact of that. Not just you made something, but what impact did it have? How many people did it help? How many stars did it get? How many shares on social, that kind of stuff? Show the impact you're having. The second point is, yeah, have a social media
impact too. So start your LinkedIn. That's a great way to start networking and allows you to connect with people beyond just your current country as well. People from England, the US, France, Germany, Japan, wherever it might be, you can start to network with these people and learn from them and they might even be inspired by what you're writing. So it goes both ways. I'd say post at least once a week there about something that you've seen that's interesting.
And if you're not found something interesting once a week, you're not looking enough because machine learning is moving so fast, you can definitely find something once a week to post if you really want to. Third thing is, if you don't get that ideal job straight away, don't be put down by that. Like try the startup lifestyle for a little bit. My personal advice would be if you're able to try startups first and then go to the larger companies later on in your career when you
want more stability. When you're younger, you have a chance at least to be a bit more unstable and try things that might just become the next big thing. Do one or two of those. And if it doesn't go nowhere, then go to the big company with all that experience you've just gained and do that because it's a lot of fun to wear men with many hats and get lots of experience early on in your career. And other than that, you're only constant in computer science is change.
So keep yourself up to date and you don't have to change like I know there's like different libraries and frameworks like in the web world as angular and there's react and all this kind of stuff. Don't worry about that too much, but just make sure you're up to date with the latest technologies that are coming out. So you know how to pair those technologies with your existing
knowledge. So I personally, I'm a pure JavaScript guy. I don't use React or Angular, but I know about the other technologies coming out like the machine learning, augmented reality and all that kind of stuff. So I can start to pair those with my core JS knowledge to then make things that other people can't and that allows you to spend standout in the crowd later on. So I think those are my three tips. Yeah, I think I can summarize basically learning is a lifelong process.
Yes. Then we'll stop learning. Second thing is marketing is also very important. How you market yourself is very important. I think a lot of people will say that all my code will speak for yourself, but it doesn't. People will be everything. You really need to put some effort and then policies up and then market yourself properly whether it's a social media or giving a talk on that with yourself. That's also very important. I think Jason, thank you
very much for your time. I think you have to, I hope that you enjoy your trip in Malaysia. Since you're from KL, I'm going to Georgetown, I'm being later. So thank you very much for your time and then hopefully we'll see you again in the near future. Awesome. Thanks for having me. Thank you. Thanks for listening. If you enjoyed this episode and you would like to help support the podcast, please share with others post about it on social media or leave a rating and review.
Follow this podcast on Apple Podcast, Google Podcast or Spotify Podcast. If you have any comments or recommendations, I would be glad to receive your voice messages. Send me your voice messages via the link in the show room. To catch all the links of episode, you can follow this show on our website, www.airmldtops.com or our social media such as Instagram or Twitter with the handler and AIMldtops. Thanks again. I will see you next time.