Introduction, roadmap and fundamentals

0:00 hey it's Andrew Brown and welcome to the free generative AI Essentials course the purpose of this course is to give you comprehensive knowledge so that you can go and start building generative AI applications and in particular this is the prerequisite course for my free gen boot camp where we actually go out and build real projects so I hope you're as excited as I am and you're able to get as as through much of this content uh so you can go build projects with me and I will see you soon okay [Music] hey everyone it's Andrew Brown and welcome to the start of our journey asking the most important question which is first which is what is the genis uh genis gen Essentials certification so this is a practical gen certification teaching you the fundamental concepts of ml AI gen all modalities of gen with a strong focus on llms programmatically working with J workloads uh both with cloud and local l M and it is a cloud vendor agnostic approach so there are other geni certifications out there it us has one Azure has one Nvidia has one and you know what I've made courses for them all

1:12 and they're okay but they're very very specific to those vendors and they leave out huge huge gaps and they have all these expectations of what you should already know in order to um uh uh to start building with Gen and so the purpose I made this gen Essentials course is so that you have round uh round and Broad knowledge so that you can be successful in any area of gen no matter the technical choice and to me that is super super important I'm going to get my head out of the way here so the course code for the exam Pro gen Essentials is exp geni 001 I do want to point out that this course I kind of feel as like a beta course at this point and so I'm hoping at some point I will release a version 002 and include all of the information that I want to actually include in this course course it's still really good comparative to other certifications it'll definitely blow it out of the water but I you know I just want to say that the release of this course will have some gaps and they will be addressed in minor updates as we receive feedback so pay attention to updates or upcoming road maps um for this content because this is not a oneandone course we're going to have to

2:20 come back to it and continuously update it because gen is moving very very quickly okay but let's continue on so who's the certification for we'll consider the examp Pro gen Essentials if you are preparing for the preit knowledge to take the free gen boot camp so you can be successful in its completion and Grading for my boot camps we go and we build we hit the ground running and so my expectation is that you can code you have General familiarity with the stuff you have uh knowledge under you so this thing is going to prepare you if you do not take this course before my free gen boot camp you're going to be struggling a little bit uh you can still do the free gen boot camp but um you if you want to maximize uh uh how much you learn please please please go through this course um and it's going to help you out

3:08 okay uh another reason you should consider taking the uh exam Pro gen Essentials if you need Broad and practical knowledge understanding gen Solutions so you have the technical flexibility to move in any technical Direction right now geni we don't know what direction it's going do you use cloud services do you do local local LMS should you be fine tuning should you be using rag um how much does it cost how do you secure it there's so many questions um that uh that that need to be answered and this is my explo exploration think of it like a CTO trying to explore and evaluate for for the company I'm coming back to you with all that information so that you can make informed decisions and have confidence building with jny so I'm hoping that's what I achieve throughout this course you need to focus on implementation deliver gen workloads that are both secure and in budget at the end of the day we got to build things for people that's why we're learning gen to apply the knowledge and so that's what I want to do I generally focus on implementation and development because I have a developer background I think it's fun building things um and so you're going to see a lot of Hands-On in this course whether you can do it or not

4:16 is a different story if you find that the the Hands-On stuff is beyond your knowledge just watch the videos and and absorb the information you can this is all about getting you exposure to the the the most amount of knowledge and and the shortest amount of time if you cannot do the things do not stress out about it okay uh let's continue on here so we're looking at the Gen road map so there's this certification course which is going to give you confidence and understanding talking about J and having the tools to get started building okay um and that is going to be based on the exam Pro gen maturity learning model which we'll look at in a separate video um the followup to this gen centrals course is the Project based boot camp which is called the free gen boot camp this will give you proof that you can build gen workloads through Hands-On projects which will be at gen.

5:03 Cloud proo bootcamp.com register if you have yet to do so go register now even if you don't finish this gen essential course you can absolutely register this is an optional course to prepare you so it's up to you what you want to do um but I I strongly recommend you go through this course if you can but still you know if you don't have the time but you do have time for the boot camp sign up for the boot camp and be there um how long does it take to study to pass this exam well we have between 15 hours and 30 hours if you are experienced you have some other geni certifications you're going to have an easy time through this but you are also going to benefit because there's going to be a lot of stuff that you didn't see if you took the adabs Gen practitioner you're going to be really surprised ads AI practitioner you're going to be really surprised how much stuff that we covered that you didn't see anywhere else same thing with the Azure AI engineer associate you'll be really surprised how much stuff we cover in this that uh I could not cover in those other ones because they're out of scope but they're not out of scope here we're covering all of gen in this course um so yeah if you uh yeah so if you are experienced 15 hours if you're a

6:11 beginner you're looking at 30 hours what is the average study time I could not tell you for this one because we don't have enough data at this point but I do feel that this is more a lecture and lab heavy practice exams are not going to be your main focus as the exam itself is not difficult it's just there to evaluate your knowledge that's why we have um this optional certification step I do want to tell you that if you want to get certified to earn this badge you do have to pay for it uh I I don't know what the cost is in this video but I know that it's not a high cost to to get it if you if you want if you are just gaining the knowledge to take the boot camp still do this course okay but if you want that extra badge to to show off you can um also do that by um paying for that that certification uh the recommended study time here is 1 to two hours a day for 15

6:58 days what does it take to pass exam watch the lectures do the Hands-On lab and again I want to point out that we are working with so many different Technologies there's going to be areas where you you might not have an account you might like okay I want to learn this but I don't really like Azure so I'll just watch the video that's totally fine or there might be uh things where you actually need Hardware to run it and so you know maybe you need a Windows machine or a Mac or a Linux machine and you don't have that thing or you need a certain amount of compute that's totally fine just watch and and do the best you can but where you can do things go do it and there's a lot of free Solutions and we try to focus as much on free Solutions as possible uh so yeah just do your best okay um do uh paid online practice exams that's just kind of something we always

7:44 throw in there for this course we have a free practice exam there's no other providers because this is a an exam made by us at this point so we will have a free practice exam so you'll have one shot to do that and then you can pay to do the uh the exam that gets you the certification okay so it's up to you that's going to be at at www. exampro doco j001 so sign up to get your free practice exam over there for grading uh the passing score is above 750 points so you need around 75% to pass you can fail by getting exactly 75% it just depends on how many um um uh questions there are because they don't always split perfectly and so you always want to aim above that score okay for response types there are 65 questions 65 scored questions you can afford to get 16 scored uh questions wrong there are no unscored questions other exams do that I don't do that in my exams I I think it's just stressful I think people would rather just know exactly what they are getting and do that there are no penalty for wrong questions the format of questions is multiple choice multiple answer case studies case studies can be

8:53 worth more than a single point so the distribution isn't perfectly like one point per question uh in case studies are more complex um um but uh you know there's not that many so I wouldn't stress out about it okay um the duration is 2 hours you get two minutes per question the exam time is 130 minutes the seat time is 160 minutes I believe that is correct yep yep yep yep um and so the the SE time refers to the amount of time that you should allocate for this exam this includes the time review instructions Show online Proctor uh Proctor your Works Space read accept NDA complete the exam provide feedback at the end where do you take the exam online from the conveni of your own home exam Pro delivers the exam via teacher seat anchor so I don't think people know this but we actually have our own Proctor system um now the question is what format will it show up uh for this

9:40 because we have an external portal and then we have one that's integrated I don't know what's set up for this I might make an additional video that's only in the platform that you'd have to watch to understand how to take the exam so just look out for that um but it shouldn't be too difficult but yeah it's all exam Pro so the place where you consume the course if if you're on YouTube and you go to exam Pro and sign up for for free you can then figure out from that portal where to take the exam and so hopefully that will make sense there uh Proctor means that it's super it has a supervisor a person who monitors the student uh students during the examination our Proctor system is a little bit different whereas uh we might capture the information and then audit it later so it's not always a live Proctor but like uh it will be it will capture information and then it'll be an audit audit period And if everything's fine then we will release the course out

10:27 to you um but yeah we we'll see how things go for that okay um this exam is valid till well it's forever because it's just how it works um other certifications what they do is they'll say good for two to three years and get recertified um it's not like we have a partner Network here so there's no forcing of uh uh recertification but if there is a major update you might want to consider getting recertified because there is just new knowledge there so I mean the certification holds for as long as there are no new versions um and that's pretty much it so yeah hopefully that is clear we didn't do an exam guide breakdown I'm going to do that in a separate video but we are going to look at that maturity model here shortly [Music] okay hey everyone it's Andrew Brown and I'm joined here with uh Rola um and so we are going to be working our way through the geni road map I believe earlier in the uh introduction I described this as a maturity model um and so originally This was um very very detailed um uh flowchart or road map that Rola had produced along with um oh

11:39 Don Don and so I I took it and I I mangled it or I should say I reorganize it into what I call a maturity model where the idea is that we can step through it and have a way of understanding where we might be in our journey in terms of a gen road map by no means is this an exhaustive list there is so much in gen but I'm hoping that um for those that are are starting to try to understand gen that this will become a a way for you to understand it and then once you have that core knowledge then you can be introduced to uh something that is a little bit more complex or robust and so what I want to do here is I want to work uh walk through with this uh with rollup and um just confirm that uh we have a good scope of learning here in terms of the

12:28 course that is that is here not everything is included in the course just because of um the strict time limit I had to produce this course but I still want everyone to see what the scope of what I think people should learn and so we'll start at the top and actually we do have a lot of content um in the course that is not in the ml introduction but I think that if you're going to learn generative AI you probably need to have some understanding of machine learning R do you think you can skip it or you or you think it's an absolute necessity to have some basics of ml I think it depends on what you want to do if you want to be just a practitioner who sets up a single system maybe you can skip it but if you want to stay in the field um You probably need some understanding of the bigger picture and you know I think with lots of things you can get away with with not knowing for uh for a while but at some point you will have to learn it and uh sometimes it's better to learn something earlier on than to uh push it back later I don't think a lot of these things are that hard like vectors and embeddings or um I mean at first they seem very

13:38 intimidating like like seeing the transform architecture and all those moving Parts but I think there's a lot of content out there that does break it down into smaller smaller parts and um uh I would just say it's like you don't you don't have to be a mathematician to to fully understand the basics of the ml um and you don't need to be a car mechanic to use a car right you if you want to rent car and use it once that's great but if you're going to use it every day if it breaks down you need to know where to look right um and also those who that are watching we're both a bit tired because yesterday was New Year's uh New Year's Eve and so I was passed up uh past past U midnight and I think Rola as well uh so just understand our energy levels are are high in our in our minds but um but anyway so after you

14:27 learn machine uh some basics of machine learning I think that's that's a good a good basis to now come into generative AI um obviously I think that a lot of a lot of people are interested in agents I think that's what's talked about a lot these days at least in what I hear when we talk about gen um so you know there are some things uh like agentic workflows and agentic coding it's a bit odd that I put them here because these are very Advanced topics but I put them in here early just because I know people want to know about them and talk about them and so it was more so to dive into these things in deep but more just to tell people yeah this is the angle of where you might want to want to go go if that makes sense um obviously responsible AI is maybe a higher priority for folks I'm not sure uh what could go in here but even though this this part looks really small basically the Gen introduction is is more so like uh I don't know like this part here um but yeah that's what I have for the geni but I think um one thing that is uh really interesting is the modalities of generative AI um and I think that a lot of folks just focus on text generation

15:36 llms but there is this whole skew of things video generation 3D Generation image generation audio generation are there any more types of generations that I'm missing out on there Rolla um I think there's a bunch on protein and DNA which I mean they can look like they're text but they're specialized types of text um there's a lot to do with protein uh folding and things like that that also been used in gen so there's a few very Niche um topics but these are the main ones I think text image video audio 3D yeah those are the main but there are Niche uh topics that have to do with health science or very specific uh Fields well it's a bit it's a bit tricky there because like you look at images and videos and and video is just a series of images or you have 3D and 3D uh or 3D files and they're just uh vertices and and text files but yeah I guess there are specialized ones and that's the thing as I don't know of anything beyond as you said like things in um the health scien Health Sciences if there's anything else that we would consider a modality but these are the main ones one thing that's kind of

16:46 really hard to figure out is like if you have video to video or uh image text to image is it the modality of text or video based on if it has one or the other the way I I put them in is like if it outputs that thing that I put it into that modality but I don't sure if we have to think about it in such clear-cut binary way of thinking these type of um uh inputs outputs for these different types of models no and I think we put them there initially to to just given people the idea that there is a lot Beyond just text generation right there's a whole slew of things happening and what usually happens is if you look at a model card if you're looking for something specific and you look at a model card they will tell you what the input and out to that model is and you look for something very specific so I think it's just enough to know that

17:34 other modalities exist both in input and output and that you can find uh a model that suits the input output combination that you are interested in and I think that since the creation of this um uh this flowchart actually there's been an explosion of video generation tools that I can't even name them all so that's why I'm kind of hesitant to put uh put names here I mean whisper is still hanging around for audio generation but uh yeah for video there's everybody has one right now um and so I don't know it just kind of shows like how fast things are moving even in a few months like there'll be just an explosion of models um in those areas but when you know the basis of these things but if you know video to video you know what video to video is right uh at least at a high level of what that means because underneath that could mean

18:21 something completely else and it's moving very fast I think all of the all of the most of the things that we've seen in the public eye are about two years years old at the moment so whenever you read an article on gen or something notice the date because things get outdated very fast well it's also really hard to figure out what to include practically because um as I'm doing stuff everyone's like oh use that did you hear of this constantly did you hear this did you hear that and I'm just opening these things and going well which is relevant and which is not but I think that um even if you do use a tool that might become dated I think that there are things that are always going to be in that realm of that kind of stuff and so it's not fully wasted information you just kind of have to move as quick as you can and you're going to have to expect that uh things are going to be uh the floor is going to

19:07 be moving uh moving around on you all the time yeah for text generation text to XX single turn multi- turn I have multimodel multimodal multimodel a lot of terms that seem very similar now we're into LM Basics which is still a generative AI a lot of stuff here we have a repeat of ethics and bias because we have resp responsible which have some overlap um and so you know looking into LMS probably the most important thing is understanding I don't know like tokenization aony of appr prompt um I'm not sure what else there is there but yeah if you fundamentally understand I think like what a context window is uh and what it can do then you have a lot of power just with llms there um we have it we're like I'm supposed to have a section on prompt engineering down here and even this I didn't do all of the videos but I realized that as was making some of the content here I subconsciously was utilizing some prompt engineering techniques that I wasn't aware of and I don't I'm not sure how much value there is to like learning every possible prompt engineer technique because there's walls and walls of of possible ways that you can do things but it is really interesting that um uh you

20:18 know if you just are you just have an idea of creatively how you can feed stuff into prompt engineering that you might implicitly be uh doing them uh how much do you keep on top of prompt engineering techniques do you feel that it's worth going learning all 200 of them or no no I don't I mean things are going to keep changing I think the the main things that have shown promise are good to know and those are usually a handful of things right and we can talk about them in the course I think understanding the anatomy of the prompt is really important that's what's going to allow you to understand um you know what you can do and and how that affects the model I think that's very important but you know um there there's going to be a lot of papers and a lot of um understanding of things as they come I

21:03 don't think they all move the needle in huge ways uh those the do you'll hear about but I think those are a handful you don't need to know all 200 no uh in terms of AI power assistance I feel like that's one of the the first entry points that a lot of people have with um uh uh generative AI but you know what when it came to these videos I didn't have much to say because they're all kind of the same um and and there's not that much functionality built around them now some some AI par assistant have some bells and whistles uh like Claude has the idea of projects or chat PT has that really cool voice feature or Google Gemini they supposedly can process larger types of PDFs and the way they do it is special and um actually I think AI 21 I don't know they have or AI 21 technically has an AI power assistant but some of these things here they're not intended to even act as AI power assistants they're just there as demonstrations so like when you go to meta or mistol I don't think you're supposed to use them as your day-to-day AI part assistance it's supp it's mostly just to show you a means to do that um but you know with with your level of knowledge do you use AI power Assistance

22:12 or or do you do you roll your own so to speak um um I tend to like to do things my way but because I trust my own systems uh I we do we do rely on anthropic CLA quite a bit uh my team uses a lot of cursor for coding and things like that uh my main issue is the the probability I would like to understand what how good what the probability confidence in any particular responses which right now okay yeah right now it doesn't tell you right it just gives you an answer know yeah I never thought about that yeah so but I guess like when you're using uh uh models direct directly would they give you probability or or would you have to do some uh some uh post calculation to get that in predictive AI most models give you some sort of probability of confidence some sort of measure of how well it thinks it did but that's that's but we were talking about just general models like any kind of ml model is that what you're saying but with LMS would they have to would they also explicitly return them as well or no so uh most llms do not because they're ating sequences word

23:29 after word they do not tell you what the what the models um how well the model thinks it did but in other probably if you're doing a classification for example it would tell you um so I guess that's a good thing it's a good thing that you said there because uh when we talk about um llms you have to everyone talks about having uh evaluations right but is it because the nature of LMS makes it hard for it to produce that number so you have to have an additional step whereas when you're using a classical machine learning model it's producing a single value and then it can also give you the probability back at the same time kind of so in predictive machine learning traditionally it was very task specific so you trained it on a very specific task and so it learned that task and it gave you some sort of metric of its performance on that result now generative AI tends to be and foundational model llms tend to be more um their general purpose so they can do a lot more things and they can do different types of things beyond the things that they were um trained to do that's where all these evaluation

24:40 metrics come in is is if you see and we talked about this at some point um is they would tell you what how does it Benchmark on coding how does it Benchmark on reasoning how does it Benchmark on accuracy um because they they they can do a lot more things than we train them on because they're very very large models and so there's emergent properties that come with that um and so I'm just going to iterate here that AI power assistance is I think where a lot of people uh start learning it but uh what happens is that when you're using AI power assistance you find out that it doesn't have a programmatic way for you to start implementing it for your own Solutions and you know the purpose of this course is to show folks how to go beyond the AI power assistance of course you can leverage this course to do the best that you can do with a powered assistance but um you really want to move Beyond in this kind of part here I almost I feel that I should have a little line here because this is where the real story starts right after oops that's not what I wanted but if I can uh get this open here it's not I don't know how to get my shapes here's my shapes but I feel that right here uh yeah if I if I was to draw a line right here this

25:51 is where things start to get serious it's obviously an arrow not a line um and this is where you start get into workbenches and playgrounds surpris techically I showed where all the workbenches were and playgrounds were in these various Solutions but I didn't really use them I just skipped right over to code because there's not much to talk about they're they look a lot like a assistance they do have more uh knobs uh knobs to turn like uh what top K and temperature and things like that once you learn one they all kind of feel very similar but uh some of them will have interfaces to their extended uh extended functionality so like if you're using a CO here they would have a little box for the response format for Json if you're in um uh anthropic uh then if you want to do tool use you could do that in place and some of them will generate at the code for you so you can start utilizing them and sometimes they gener out good code I can't say they always do I think in some of them they're actually producing older code I'm actually really surprised um with how many AI Services uh like they'll have a version to uh API and and then their main service that Jour toota is still doing version one uh and it almost kind of

27:03 feels like uh these companies kind of even keep up with their own their own changes which is really surprising um but I'm not sure what your experience is about um documentation and how how accurate it reflects the current versions of things usually they're they're behind I think part of it is the the speed of things but also I think part of it is cost right the the older mod model is likely smaller and it's likely cheaper uh and so some some of it is cost um so you know once you start working with things programmatically at least in this way before you even start writing code you might start to think about what kind of models you want to use and there's all kinds of models out there there's foundational models we have models that are fine- tune which technically Foundation models are fine tune models models that are instruct models we have embed models we have open source models we have Edge optimized models or we could call them slm small language models or we could call them on device models or we could call them something else there's a lot of names for this one in particular um and then we have uncensored or you could say unrestricted I don't know the word

28:11 uncensored doesn't sound good but unrestricted models uh and we have a lot of models why do we have so many darn models um so I think these are all different terminologies to um to describe different metadata about the models so open you know open source models are models where the weights are open um plus the data plus the code open weight models are models where the weights are open but not the data and the code so each word descri each term describes a very specific scenario so Edge optimized models are usually smaller models that you could put on your phones and Edge devices as opposed to the larger models that need a server so because there's so many different um ways that you could use these things you have to have words to describe the different situations I guess and I think there's probably a lot more open weight models than there are open source models um there might be a bit of a confusion between that because I always think like I like this is open source model like but is it I mean I can download the weights and I can do things with it but I don't really know how

29:20 they've trained it um one model mistol and um uh meta tend to be open weight models but they're not open source models so the data and the code that generated the models are not there um open Ai and uh is is a closed model and anthropic is a closed model um yeah so there's different ways in which these things are released allowing people different types of control um and that's where the the naming comes in and and I like if people are watching here yes I'm updating the this as we go because that's how it is you collect information and you you find new ways of organizing it uh I know of of one open source model that I can that I can always remember which is IBM Granite because they say that their model is truly open source and and by that they mean like you even know what data was sourced to train the model um so it's really interesting the different levels of things there we have models of service so I think this is probably one of the easiest ways to start working with llms um and and some other modalities uh where they provide a a

30:35 unified API to access to multiple different models um and so Google Amazon Azure and Alibaba all have them actually I haven't tried Alibaba yet I just know that they have it and I was I was I I said I was goingon to go use Alibaba I never did um is there any other ones that are out there the only other ones that I can think of that are models of service is grock grq for those that are listening there's a lot of GRS um in the AI space and it gets a bit confusing but that one was amazing however um they don't have a paid a paid tier so like you can use it and it's super fast and it's only for open source models but um uh they don't have a a payment plan so you can't even utilize for production they do have a business plan but um it's a bit hard to get in there and so once we're done with models of service now it's time to get serious and we're actually looking at um out like some of these are cloud services but now we're getting into a very serious um development and this could involve actually downloading models to your computer or uh using uh best-in-class thirdparty cloud services

31:46 and so this is not exactly organized the same way as the course but now we are into Dev tools and workflows and so something that folks probably should know about Is the Gen life cycle l mops these are Concepts that sound very similar in the cloud space when we talk about I don't know the development life cycle and devops but um they can be applied there as well and then there's just some core developer tools that I think that people are always going to come across like I'm on hugging face hugging face all the time it's like the GitHub of the AIML world I was surprised that I I used AMA more than I thought um uh do you ever touch Ama or is it um it's not for you Rola I have but I I know people love it it's it's open source there's a lot it's a lot more open than a lot of other systems so people like it I haven't personally played too much with it but I I don't even remember I don't even remember what LMS is I'm not even sure what that is but it's in there uh so I can't remember that must have been something at one point but now I don't know what it is almost tempted to delete it we got llama

32:53 file Lang chain llama index these two things were things that I heard a lot about early on when I was I was learning about generative Ai and I actually ended up once I started gaining so much knowledge I don't even touch them at all now I don't even think I even bothered to teach them because even though they were uh I think they're actually really good as a as a learning tool um I just found that uh the the promise of what they were supposed to provide was to give you an opinionated framework so you could rapidly prototype things but uh because this space again has mve so quickly um a lot of the documentation was out of dat or The implementations Underneath didn't really live up to the promise um do you use any kind of um orchestrational uh services or uh Frameworks for rapidly building out llms or or like agents or workloads or what

33:41 do you do so our team usually uses sometimes Lang chain um and guardio to create very fast prototypes um but like you said everything's catching up so even Bedrock we use a lot of AWS bedrock and now they can you can do a lot of these things natively in Bedrock now you can do a lot of workflows and and model changings and so on which uh you know they they were quite you know some of the first tools that came out but now a lot of things are catching up like you said well yeah and and it's it's really surprising like how robust Amazon bedrocks coming on the Azure side they were they were leveraging open AI but now they have their own generic one like Amazon Bedrock um I don't not sure why I put gml over here I mean they it's by the same creator gml and ggf um but uh llama CPP that's something I just want to do more with because it sounds so awesome but um and it comes up a lot but I I I didn't get to spend too much time there with it um and Onyx is something I again I I really wish I had time to talk about that more because that came up a lot um for compatibility for model weights uh I try to remember

34:53 what it does it's it's a file format it's like is it like ggf but a different file for you remember what Onyx Onyx is in particular no I'm not sure actually so this one here um I remember it's a this is for interop interoperability um and that's the idea of um kind of using the same formats or the same um the way you encode the the models and all of that in the same way so that you can actually exchange use the same systems across different models and across different providers and so on um I know a lot about interoperability from healthcare because everybody does what they want and it's a becomes a headache which is where the space in AI is as well right every provider um releases in whatever encoding they want and tokenizer they want and they encode on disk the way they want and so that uh makes it a bit hard to um kind of use the similar systems across models and across so that that's

36:04 great yeah it's definitely an interesting topping it it would make life easier for everybody if we have certain uh Common Grounds see see I knew you knew it because as soon as I opened it up it took you two seconds to know what it was um app prototyping i' I like I come from a developer background so for me I I really enjoy AI code code assistance and app prototyping and this course we covered Streamlight gradio I think we did Fast HTML uh vzero and lovable there's actually a few other ones that that I had covered as well um and then on the other side we have ai code assistance so we Amazon Q developer I I did cover that one GitHub co-pilot Gemini code assist codium cursor um and these things are actually this one has codium in Wind surf now which uh I would say I really like wind surf but again you know things change every other day um and uh it's just neck and neck these these tools are changing and so having general knowledge about the tools and not really being focused on just one thing with conception understanding is going to just make your technical path a lot easier because you're not never going to be left behind uh and you

37:12 basically are just jumping ship constantly to new tools but it's not that big of a deal except when every tool cost $20 then then you have to decide decide which one you're going to unsubscribe to I put extra emphasis in this course on setting up the developer environment I found that this was um one of the largest challenges with working with local models understanding how to set up cond uh installing Docker containers uh pulling containers uh setting up Jupiter a Jupiter server because that's a very com a common way um and I was actually really surprised on how many free options there are right now to learn I think a lot of people think that uh well it is expensive to run models but a lot of people think that it's that it's too prohibitively expensive to learn gen but I is just blown away by how many resources are right now um as every company is putting out their own Kool-Aid for you to drink that they really want you to uh buy into it and I'm sure at some point that this will change as as the um uh the space matures and everybody understands that but right now is like the best time to learn because there's so many free free resources it's

38:22 unbelievable um but uh would you do you agree or would you say do you think do you see a lot of a lot of free learning opportunities right now or like with I do I think there's a lot from from all the big providers AWS gcp uh hugging face has a lot of things there's a lot from a lot of different resources all small startups as well I think there's always great resources um it's a matter of time I think and commitment well even like in hugging face they have all these projects running here and you open it up and they have something called like um uh something zero it's and I don't know what it is but it's it's like some kind of compute that is distributed that allows you to run anything this is not exact the one I meant to click on but it's just very interesting that there are so many of these things running and I'm just like how do they how do they have the money for all these for all this compute but they do and it's not showing the name of it's called g0 or something or um something but um uh yeah it's just it's just really interesting one part that's a little bit hard for me to cover is ji security I'm not sure if it's because things are still emerging we do have a few friends on the security

39:32 side that are going to help us out at some point to um expand on this but I imagine that this is uh rapidly changing to match uh the needs of of this expanding Market um of course learning how to work with containers I think is very very valuable there's a few different solutions not just open AI Opa but uh Nvidia has one called Nvidia workbench um I I would just say like there's not I haven't seen much tooling around containers and k8s and when I've talked to folks in the kubernetes uh or Cloud native Community they're all like I know containers really well but I don't know anything about J and I'm like but you know containers right um and I think that um those that know how to work with containers that uh they they might have um uh an extra Edge if you will for deploying things into production and we definitely need people that know how to do that so serving models I think um uh is is also very useful it's it's unbelievable how many ways you can serve a model just as web servers have servers uh llms have servers and I also assume that um um General models have servers um but I here's a a small list of things

40:46 in fact I covered I don't think I covered P LM but there was a few other ones that I I covered um do you have to deal with uh serving models uh very often or uh uh what what does serving models look like for you Rola yeah so when you when you put a model out for usage people have to we call it inference or survey you have to put it on a machine for people to Ping that machine and and get their answers and but we tend to do a lot of things native to AWS so you can do that through Sage maker and Bedrock quite nicely so um but yeah if if you want to put it off the out of the cloud or use your own Solutions then yes but the main point is to actually serve the model right that's how you make it public and and usable yeah and I think I think with all major providers with with their specific uh ml platforms that they have their own their own uh inference engines or whatever you want to call them and I remember the one for sage maker uh yeah these ones these ones are really interesting as well Ray I I really stood out to me because it allows you to do things distributed um so I didn't realize that a lot of these things were

41:53 um that only serve on a single machine and you have to in do something like Ray to take BLM and then spread it across machines which was really interesting uh there's model optimization um so I have a few here a few things here listed um stuff keeps shifting around here there's another one here I don't know think I covered it but again merging stuff but it was called like webn or something webn or something I can't remember what was called but it was a Microsoft project um and so that stuff's changing all the time but one thing that I I thought that might be really important is understanding the underlying uh uh Hardware um I think that um we really don't know what direction uh the Gen at least I don't know maybe R you know but the the the direction of of what it looks like to build a workload right and and what do you need to do that like uh because I know there's different levels of of models and inference and so I in my mind I think of it as like lightweight medium and heavyduty uh inference and training and so when I say that I mean like if I have a a computer like a modern computer and it can run a uh let's say a seven

43:05 billion parameter model uh that's like a light that could be a light workload something that that is feasible that I could run or something that could end up on my phone one day and so I think okay what does that do for the end end business or customer if now they can run this size of model locally right like and you and you have this one Capital cost of like a TH or $2,000 machine and now you can build a workload that's that's scoped only for your office and is that a benefit to anybody and so because the stuff becomes is becoming more and more possible to run on uh on our devices and stuff like that I keep thinking well we need to know a little bit more about this right um uh but I don't know if you have any thoughts about uh the hardware side of it or I mean the hardware side is interesting because these models are so large that they push the limit of what we know about Hardware right so we can we're going to talk a little bit about model sizes and what that means for running them and and so understanding how these models have pushed our need for custommade chips and things like

44:13 that is fairly interesting and um and like you said having Edge um compatible or Edge optimized models can change applications in in our world as we know it and I think we've seen we saw a lot of different phones this year with their main catch being that AI now runs on your phone right so Edge um gen on edge is is predicted to be a very very lucrative and big Trend well it it'll be very interesting to to to see what happens and just the other thing is just understanding how much resources these things take because when you see or you can uh kind of think about how much it takes takes and then you might ask like how is this sustainable and what's going to happen to some of these uh services that are out there and how are they going to make a profit um and you know if I do invest my time and effort should I go with this or should I go with something that is more feasible that I can run on at this size and so uh it's really hard to say I was talking to one of my friends and I'm sure they're going to be really interested in the boot camp um because they're that's that's they're really

45:21 excited for the boot camp and specifically the hardware side and and they wanted to go and spend a bunch of money because they wanted to build a home lab because they don't want to get the the the rug pulled on them so to speak that the price is going to go really high and now they have this uh service that becomes like essential for their for their dailies and so um they're trying to invest in that so I thought that that was an important inclusion of it whether I did a good job on that is is up to the folks that are watching but hopefully uh it is good then we're getting into more advanced territory and basically everything below this line I mean Onyx probably is below this line too but everything below this line is more I would describe as Advanced Techniques um though to be fair if you're going to be deploying models you're probably going to want to have evaluations and guard rails and things like that but I don't know I think all

46:06 this stuff is um and maybe this is the stuff that I didn't get to cover much in detail this is probably where um I've done the the the poorest job in in this this course but maybe that should be its own course because all these things are uh very involved at least I think they are um but uh do you think that uh folks that are becoming Engineers or or maybe not AI Engineers but they're picking up these generative AI skills for their companies do they need to learn super fine tuning like know how to do it or just uh like actually be able to do it or just know of it um I think knowing of it is always a good thing I think it depends on the field you're in if you are in a general field then these models are advancing fast enough that you might not have to do it on your own but if you're in a uh

46:54 Niche field where the model needs to be fine-tuned on your specific field of data then probably you need to know but if you're in a general company general use case scenario just knowing it is exists is probably enough at this point I think the only thing that might be missing from this is all the general data skills that you need and um uh I guess I just maybe it needs its own data primer I think I might have put some data in the course I think I had some data data content there but I don't have any data specific video saying like this is how you prepare data for for ML models or llm models um how important is data to your models does it matter the data that you put in your models you're asking me how important data is forl models is that your question I know is it like well I'm just saying like well I'm just saying like imagine you wanted you have something you're like oh I want to train a model and I wanted to do something is 20 lines good enough like how much how much time do you need to spend on data comparing comparative to uh comparative to actually um engineering llm

48:04 Solutions so um I think let's start with regular uh ml before llm data is the bread and butter of machine learning right if the model learns from data um and then and so the data has to be there it has to be of good enough quality and it has to be um pretty consistent then and so that the model learns um from it now in terms of llm specifically it's less of an issue on you it's not an it is an issue for the model of course because it's still learning it's less of an issue on you because the model provider of that foundational model has done all the work and unless you need to fine-tune it or change anything about it or make it more uh you know for your field you don't need to worry about it right because it fine-tuning llms is more a privilege than um not everybody can do it uh how much data are these uh things trained on well in the in the words of Yun who is one of the pioneers of uh deep learning and machine learning

49:18 um it take 20,000 years for me and you to read That's how much data these things get to see before they become useful uh so can you and I sit down and curate a data set tomorrow maybe if we commit a few months to it but it's not it's not for the faint of heart I I I kind of Wonder like with data sets at large how do they know what's good and bad data when they have to consume that much data or I guess it just kind of evens out eh no well that's the thing is is we're seeing all sorts of things and we see and and we see biases a lot of the data is also scraped off the internet right there's biases in them we're seeing these models being biased they they are perpetuating social bias because these days the sets are not necessarily cleaned it's it's also really important to know what data sets the models have been trained on I think I was looking at some of the Google models I can't remember which one but it's something to the tune of 40% of the data is from social media right like that's thing that's not good that's not good right there not good at all curate

50:27 social media data there's nobody who there's biases in social media data there's no accuracy so what data and and that's the important part for example between open source versus open weight models is that there's more transparency over the data that the model has seen um and so if you have a chance to look at what percentage of uh social media data what percentage of literature what percentage of of Science and like whatever the breakdown of the data content that the model has seen will give you a better understanding of that what what I what that model looks like right uh it's like it's like looking at somebody and saying did this person go to college or they they spend 10 years on social media that could be a model right depending on what sources it saw yeah I I just think of the movie uh and for those that have not seen it uh you might want to see it there's a movie called idiocracy uh where um it's oh R do you know the movie my husband likes that movie there's a there's a really good line in there I always forget it's uh it's Bravo or bra it's the the quench you anyway it's the product that's in

51:39 the the movie but uh yeah it's a bit scary to think like that they're scraping social media data but even even um again I think there's a lot of data it's just it's just it's interesting to me but anyway I just wanted to show people the road map to give an idea of of the scope of it um hopefully that gives uh people an idea of the boundaries or or scope of Genera AI this stuff can be a lot more complicated so I did my best to um simplify it and break it down um but uh in reality if this was if this was graphed properly things would be all interconnected all over the place um and and and I made it look like a journey through one step whereas it's really all over the place uh or if anyone's ever played the the video game p the Exile um which has like a skill

52:26 tree um there's the pth to Exile skill tree I feel like it would look more more like something like that if I click into here and it's just all over the place that's kind of what geni um yeah here it is you go here and it's just it goes everywhere right and so I feel that if we did it like that it would look like that and people would be like I'm done but um I think that uh this is a a fair compromise here uh thank you Rola for helping us here in the road map uh for folks that are watching Rola is going to show up in a few other videos here uh and um yeah that's about it CIA [Music] Chow hey this is angrew brown I just wanted to point out that there are some areas that I wasn't able to fully record um just because of time constraints but also there's just other really good resources out there that you can utilize um to fill any gaps that you need before the Boot Camp starts or to uh make sure that you're successful if you're trying to get the certification for the uh the Gen Essentials course uh so I've compiled a list along with free camp at gen.

53:12 cloudro bootcamp.com boo forre Camp of a bunch of resources and we probably will keep adding to this list of things that you could utilize um uh uh utilize to fill in those gaps so for example fine tuning was something I didn't get to do a whole lot about if you go over to here you can see that there are ones on fine tuning right uh or uh you know Google's offering I didn't get to do a whole lot about but there's tons of things about here with the Google offering some of these are my videos as well and so you know I'm just suggesting you to uh take whatever you can to prepare um and that list is here at the gen.

54:00 cloudro bootcamp.com boooth slf freeco Camp so there you [Music] go hey this is Andrew Brown and we are taking a look at the definition of what is artificial intelligence and we really wanted to put this against uh the terms of machine learning deep learning and generative AI so that it's very clear what the differences are often people just say AI when they mean ml or deep learning so understand that um these terms are uh not used correctly often but people generally will understand what you're trying to say so it's not a big deal if you use them out of turn but let's make sure that we know what they are let's so let's first take a look here at artificial intelligence also known as AI these are machines that perform jobs that mimic human behavior okay that's the key thing here is that they are human like or doing tasks that you'd expect a human to do um and that is clearly a very broad term of what is

55:11 AI and so you can see why a lot of things are attributed to being AI then you have machine learning and machine learning initialized as ml is machines that get better at a task without explicit programming now now of course we have to code a machine learning model but once we have that model and we pass things into it it's able to complete its task with its very complex algorithms um so you could also just think of it as an uh it's a a a special algorithm to perform a task that would negate you the negate you having to do calculations or programming or things like that then we have what is deep learning and when we think of a lot of the AI stuff we're usually think of deep learning because it's these machines that have an artificial neural network inspired by the human brain to solve complex problems so you probably have this uh you probably seen a graphic of it of like these nodes and they're interconnected and they go through layers that's deep learning a lot of people call that machine learning or AI but no that's that's the L then we have gen so gen which is more of a u marketing term but generative AI is a

56:23 specialized subset of AI that generates out uh content such as images video text and audio now I don't have it in the graphic on the left because it's hard to say where does it go does it go here right because it is a subset of AI but technically um gen often utilizes deep learning because when we think of it and my Line's not dry here today but um when we think of it there we go there's the line is that a lot of gen techniques like large language models or um or um Vision models things like that are utilizing neural networks so it is deep learning [Music] okay all right so I know we keep talking about what is AI what is Gen AI but we're going to cover it again just so that it becomes more clear from different perspectives um so let's talk about what is artificial intelligence so AI is computer systems that perform tasks typically requiring human intelligence um so these include things like problem solving decision making understanding natural language recognizing speech and images and an

57:32 ai's goal is to interpret analyze and respond to human actions it's there to simulate human intelligence in machines when we use the word simulate we're talking about mimics aspects resembles behaviors but what we're not talking about is emulation which is replicating exact processes and mechanisms as if you created literally a virtual human brain that's what emulation would be um so AI applications are vast and include areas such as expert systems natural language processing also known as NLP speech recognition robotics uh and more AI is using various Industries for tasks such as uh we're talking about business to Consumer so think of a customer service chatbot if we're looking at e-commerce think of a recommendation system if we're talking about the Auto industry uh maybe we're looking at a atomous Vehicles if it's medical then medical diagnosis there's a lot of applications for AI but it's a broad application for all sorts of things now let's take a look at generative AI so generative AI uh often initialized as geni or or said as geni is a subset of AI that focuses

58:47 on creating new content or data that is novel and realistic it can interpret or analyze data but also generate new data itself it often uh yeah so like types of content produces would be text images music speech and other forms of media it often involves Advanced machine learning techniques uh so it could be using things like Gans it could be using vae so variational Auto encoders um a lot of current llms use the Transformer architecture so if you're using um chat GPT or Claud Sonet or any of the popular ones they're basically all Transformer architectures uh gener I has multiple modalities and when we say modalities it's like think about your your senses you have touch taste hearing smell so modalities are the kinds of content or or um senses that a model has so we have Vision so realistic images and videos text generating humanlike text audio composing music molecular which is more of an interesting one so drug Discovery via geomic data and uh I want to make it clear again we're talking about large language models but llms large language models will generate out humanlike text

1:00:05 and is a subset of gen it's just one modality of the many modalities um but it's often conflated as being AI or gen AI just because it's the most popular and In Demand right now and the most developed so just make sure that you understand that gen and AI is not all about large language model it's just one modality one application of of the broad sense of AI and gen now let's just make sure we have a side by-side comparison uh and then I'm sure after this you'll definitely know uh definitively the difference between Ai and geni so in terms of functionality AI focuses on understanding and decision making whereas geni is about creating new and original outputs for data handling AI analyzes and makes decisions based on existing data geni uses existing data to generate new and unseen outputs in terms of applications AI spans across various sectors including data analysis automation NLP and Healthcare where gen and yes I see the spelling mistake uh it's creative and Innovative focusing on content creation synthetic data

1:01:17 generation defix and design so there you [Music] go let's talk talk about Jupiter so Jupiter notebook is a web-based application for authoring documents that combin live code narrative text equations and visualizations and before it was called jupyter notebook it was known as I IPython notebook and jupyter notebooks were overhauled and then turned into an ID called Jupiter lab which we'll talk about here in a moment but you generally want to open notebooks in Labs um and the leg the Legacy web-based interface is known as Jupiter classic notebook and to be honest I get confused between Jupiter lab and classic I think most things that use these days are Jupiter lab um but the confusion is because we just call them notebooks even though Jupiter classic notebook is the not note the uh the older one and the newer one is Jupiter Labs let's go take a look at jupyter Labs so jupyter lab is the next Generation webbased user interface it has all the similar features as the classic jup notebook in a flexible and more powerful use interface so it has

1:02:27 notebooks terminals text editor file browser uh Rich outputs and the way you I think that you know that you're using Jupiter lab is that will have this uh these tabs here on the side and a bunch of functionality so Jupiter lab will eventually replace the classic jupyter notebook and that's kind of true because um but not fully because in some places I do come across classic notebooks launching them up um but for the most part functionally it has been replaced then we have Jupiter Hub so jupyter Hub is a server to run jupyter labs for multiple users it's intended for a class of students a corporate data science group scientific research groups and so it has some components uh underneath you will come across notebook like experiences that are like Jupiter Labs so some companies will um extend the functionality of it one example is Sage maker uh Studio Classic for whatever reason adab us um spent all this time creating extensions and extending Jupiter lab and then they decided uh no we're not going to have extensions anymore and we're just going to use the vanilla version um but uh there's also things like vs code that has notebooks

1:03:38 or code lab that have notebooks and vs code is like its own kind of notebook thing it's not juper Labs but it's jupit lab compatible so just understand that you'll come across things that are notebooks that look like Jupiter lab but they're not necessarily Jupiter lab okay [Music] let's take a look at natural language processing also known as NLP and in machine learning it's a technique that can understand the context of a corpus a corpus is a body of related text the text that you are working with and NLP intersects with computer science and Linguistics so if you know a lot about the the nature of uh spoken and written language then uh computer science here is going to meet in the middle here so that we can um make sense of it using algorithms so NLP enables us to do things like analyze and interpret text within documents emails and messages interpret or contextualize spoken texts like sentiment analysis synthesize speech uh such as using a voice assistant talking to you automatically translate spoken or written phrases and

1:04:47 sentences between languages in uh interpret spoken or written commands and determine appropriate actions another thing you'll hear a lot is language understanding which is supposed to be it's a it's more like a specialized subset of NLP um uh that just goes farther to understand uh more traditional older ways of doing NLP but uh anyway what I'll do is we'll just take a look at this um very simple flowchart to give you some idea of things that are related with an NLP this is mostly just to get you exposed to some terms it's not important to remember what these are and I can't even describe them off the top of my head um but again just you exposure to NLP terms so that when you see them later you'll go look up and be like I remember seeing that term here so here we have like text wrangling pre-processing language understanding so structure and syntax processing functionality which is what the NLP uh does for you in the end but text text R pre-processing is where you are preparing uh text to be uh put into possibly um a machine learning model or maybe you're using it for um some kind of analysis or something like that and so this is basically taking text and um

1:05:58 formatting it changing it and so what could we be doing here well we could be doing conversions maybe we're lower casing things maybe we're upper casing things um maybe we're turning contractions into their full forms or vice versa sanitation this is where you are maybe stripping out HTML or special characters or you are removing stop wordss when uh you have stop wordss later on in your ml models tokenization which is converting um the text into uh Vector embeddings we have stemming okay we have uh lonization so there's a lot of things here but you can see it's mostly just like formatting the text to be utilized for something else we have language understanding so these are processes to make sense of the text so part of speech tagging so is this an adjective is this a noun things like that chunking how can we uh break up the text and then work with those chunks later on down the road so that it still makes sense dependency parsing so you know which word relies on other words and what relationships do they have to other ones uh consti consu parsing very hard for word for me to say but like imagine a um a a

1:07:10 tra tra green and so like you know a noun has an adjective under it which has another thing under it you look up if you look it up and go to Google Images you'll you'll know what I'm talking about then we have processing functionality what are we using NLP 4 so we have name entity recognition this is is where you have a body of text and it's highlighting uh important words like maybe important nouns that it thinks you you care about or things like that or personally identifiable information we' got engrams sentiment analysis is this text positive negative happy sad information extraction what are we trying to get out of a large body of text um same thing with information retrieval questioning and answering topic modeling so you know again not super important to know these in depth right now but the things that are important we will see these terms again um and you'll know what they are then so don't worry about trying to memorize this now but just get that exposure to NLP terms [Music] okay hey this is angrew brown and we're looking at the concept of a regression and this is a process of finding a function to correlate a label data set into a continuous variable or number uh

1:08:18 so imagine we need to predict a variable in the future such as the weather what is it going to be next week and so the idea is that you're going to plot your data onto a graph or vector space our dots are represented as vectors um and we're going to draw a line through it which we call a regression line and the point of the regression line is that is our prediction so if this is going over time based on the temperature um you know uh that is how we are figuring out in the future what things are going to be so the distance of a vector from the regression line going to just get out a different colored pen tool other than red so maybe s and so imagine this dot here to the line that's what we're going to call an error because the idea is that um things that are closer to the line is the prediction and things that

1:09:05 are farther away from the line are an error from the line so hopefully that makes sense there are different regression algorithms used uh uh that can uh that we use to predict future variables so we have mean squared error uh root mean squ error mean absolute error and So based on the algorithm that you use to draw your line that's going to change um the prediction [Music] okay let's take a look at classification this is the process of finding a function to divide a label data set into classes or categories so the idea here is we're going to predict a category to apply to the inputed data so will it rain next Saturday is it going to be sunny or is it going to be raining so the idea is we have our data we're plotting it on a graph but we're drawing a classification line that divides the data set okay and the idea is that if it falls on one side then it's sunny it falls on the other side then it's rainy and so again if you have a different type of algorithm that's the thing that's doing the division um it's going to have different results you have a a

1:10:17 logistic regression a decision tree random Forest you can use a neural network you can use a a a name AV v b I always say that wrong so I do apologize or you can use KNN or you can use a support Vector machine at or svm so just understand that there could be more algorithms of this but these are the common ones and you know if you want to learn more about how these different algorithms will change just look up on the Internet uh what that would look like and there's definitely visualizations out there [Music] okay let's talk about clustering this is the process of grouping unlabeled data based on similarities and differences the key word here is unlabeled when we looked at uh um classification that was labeled data so the idea here is that we're grouping based on similar user differences so imagine that this grouping of dots that are close together we determined that that is Windows and this uh group of dots are Mac computers and just like classification progression you have different algorithms are going

1:11:25 to give you different results and the reason why I show you these algorithm names is because when you have to do classification regression or uh clustering uh you're going to see these names because you're have to choose what algorithm you want to utilize right now it's not so important to uh know them but when they are important we will look at them uh in more detail [Music] okay so we are going to dive into the types of machine learning in other slides in more detail but this is just kind of an overview so that you can kind of see these terms up front um so we'll just quickly go through this here and we're going to group them um based on what they're trying to do so the first is learning problems we have supervised unsupervised reinforcement these are three terms you're going to hear quite a bit with machine learning uh the key thing here is that supervised is where you have labeled data and unsupervised is where you're working with unlabel data for reinforcement this is an agent an agent that operates in an environment and must learn to operate using feedback and this kind of sounds like agentic workflows or agentic coding we're talking about gen which we'll learn about later but the idea is like imagine you wanted to make a uh a machine learning model that

1:12:37 played the the Mario or or the Sonic video game that'd be using reinforcement learning okay then we have hybrid learning problems so we have semisupervised self-supervised multi- instance so semi-supervised is where you have a mix of labeled unlabeled data you have a lot of unlabeled data and a little bit of labeled data and so um that's kind of a a mix between supervised and unsupervised you have self supervised um and I believe that this is where um the idea is that it can label its own data I think but we'll find out later on in future slides we have multi- instance where we have um examples of unlabeled data and so then we just kind of bag them together um again we'll cover that later on we have statistical inference so here we have inductive deductive and and transductive so using evidence to determine the outcome or then we have deductive using general rules to determine the specific outcomes and then we have transductive used uh to predict specific examples given specific uh specific things from a specific domain okay then for learning techniques we have multitask active

1:13:47 online transfer and Ensemble so multitask is fitting a model on one data set that addresses multiple Rel problems active is the model is able to query a human operator during the learning process um online is using available data and updating the model before prediction is made kind of sounds like rag when we're talking about gen um but again this is just general machine learning right so we have transfer and model is first trained on one task and then some are all the models used as a starting point for uh for related task and then we have uh Ensemble where uh two or more models are fit on the same data and the predictions from each model are combined so yeah we're going to see these terms again but just trying to get it uh up front here for you [Music] okay let's take a look at the divisions of machine learning this is just another way to break up machine learning um and these terms you're going to see uh more in how we're going to structure our upcoming slides here so I just want to give you a quick overview here so we have classical machine learning and the advantage of classical machine learning is the data is simple you have clear features um um and generally classic machine learning is extremely uh cost

1:14:59 efficient compared to other types of machine learning but this is where you have supervised unsupervised uh kind of uh stuff so you know when you think of classical machine learning think of those two things supervised and unsupervised um uh learning then you have reinforcement learning this is uh when there is no data and the idea is that the model is going to through trial and error figure out what is the right thing to do this is where we have real-time decision- making game AI so we talked about Mario or sonic uh uh like the the ml model playing those games and failing again and again and again until it can pass the game a learning task or robot navigation so think of automous uh driving vehicles that would be a good case for reinforcement learning we have Ensemble methods when uh quality of data is a problem so then you are going to have different strategies to work with multiple models or algorithms to have a better outcome and here we have things like bagging boosting stacking okay and so you know you'll see those terms like boosting you'll definitely see the word boost more uh when we get to that then we have neural networks and deep learning you should just really think of

1:16:07 deep learning as neural networks this is when the data is complicated and or the features are unclear this is where you'd use uh neural networks like a convolutional neural network a recurrent neural network uh a gan so generative adversarial [Music] adversarial Network sorry U multi-layer percepton uh or perceptrons sorry MLP Auto encoders and I just have a really hard time pronouncing these things but yeah you're going to see these terms again so again don't worry about it right [Music] now let's take a look here at classical machine learning and so when we say classical we're talking about algorithms that have existed for quite a while may maybe as early as the 1950s because we had these mathematicians and they figured these out and a lot of these things actually relate to um statistics right so we're taking statistics um and utilizing them uh in these algorithms in our Computing spaces so hopefully that makes sense but yeah it's they're called classical ml because we are dealing with algorithms and one example would be nearest neighbor algorithm which was invented in 1967 and lots of companies today

1:17:20 definitely could utilize classical machine learning uh to solve business problems just because they're old does not mean that they're not good it's just a matter of organizations knowing how to adopt uh classical machine learning so let's talk about first supervised learning so this is where we have data that has been labeled into categories and this is great when we are doing something that is Task driven we're trying to make a prediction because the idea is we have this labeled data and so then we can bring unlabeled data and tell the machine to label it right so here we have classification so imagine we want an outcome this would be to predict the C what category something belongs to a use case here would be identity fraud detection we have regression this is where maybe we want to predict a variable in the future so we're we're trying to figure out a market forecast um and we cover you know classical regression so you should know what these are um if not you will know about what they are soon enough because we'll cover them more than once um then for unsupervised learning we have data that has been not been labeled okay this is where things are data driven so we recognize a structure or a pattern we're

1:18:29 not making a very specific prediction um here we have clustering so the outcome of something so you group data based on similarities or differences example here would be targeted marketing Association so find a relationship between variables through Association the use case here would be a custo a customer recommendation we have dimens dimensionality reduction so here help reduce the amount of data or pre-processing this is a problem you have a lot of data um and this a use case here would be big data visualization so um yeah there you [Music] go all right let's compare supervised versus unsupervised learning and I know we've already talked about it like twice before but we're going to talk about it again and then again because I'm just trying to give it to you in different perspectives so that you really know the difference between these so let's talk about what is supervised learning so this is a machine learning task or function that needs to be provided training data and the training data is when you provide labeled data the correct answers and the Machine can learn from those results so show me how to do it and then I can do it on my own

1:19:38 that's what's happening here and so for supervised learning models we have classification regression what about unsupervised learning this is a machine learning Tas or function that needs no existing training data uh for this it will take the unlabeled data and discover its patterns applying its own labels so I am an independent worker I can figure this out on my own right uh and for this these unsupervised learning models we really should have put the unon that there let me just fix that there unsupervised we have clustering Association Dimension dimensionality reduction and so supervised learning tends to be more accurate than unsupervised learning but requires more upfront work whereas unsupervised learning still requires human intervention to validate the results so hopefully that is clear [Music] okay okay let's review it one more time I know it's getting tiresome but it's very important that you remember the difference between supervis unsupervised and reinforcement so supervised learning is where the data has been labeled for training it's Tas driven and you're making a prediction this is when the

1:20:45 labels are known and you want a precise outcome when you need a specific value return and so here we use classification regression as example examples of supervised learning there's more than just those two but that's what I want you to know for now we have unsupervised learning data that has not been labeled the ml model needs to do its own labeling it is Data driven you're recognizing a structure or a pattern when the labels are not known and the outcome does not need to be precise when you're trying to make sense of data here we have clustering dimensionality reduction Association then you have reinforcement learning so there's no data and there's an environment and an ml model generates data and many attempts to reach the goal this is decision driven you have game AI learning task robot navigation so hopefully that is clear and it's in your head um we're going to repeat these again but it's going to be less of this um and more detail [Music] okay let's talk about supervised learning models and we're going to cover classification and regression again um just so that we really know that we

1:21:57 know what these things are so classification is a process of finding a function to divide a data set into classes or categories so imagine will it be cold or will it be hot tomorrow right so very clear it's either one or the other it's going to fall on one side of the line or the other one we have different algorithms we can use like log Logistics regression K nearest neighbor support Vector machines kernel spms uh Navy's uh bay decision stre classification random Force classification so we're listing a lot more here we have what is regression regression is a process of finding a function to correlate a data set into a continuous variable number so what is the temperature going to be tomorrow and here we have uh things like s uh simple linear regression multiple linear regression polom regression support Vector regression decision tree regressions random for regression just again want to continuously repeat that so you know what these things are [Music] okay let's take a look at unsupervised learning uh so what can we do here we

1:23:11 have clustering and again we've covered these prior but I just really want to make sure that you know what they are so clustering is a process of grouping unlabeled data based on S similarities and differences right so we Ed an example previous viously um you know is this a Mac or is it a Windows here it's about age and something else and so it's saying you know is are these people do these people have cholesterol are they highrisk or lowrisk um for chering algorithms we have K means uh DB scan K modes then we have Association so Association is the process of finding relationship between variables through Association um so the idea is that if somebody buys breads then suggest butter because based on previous combinations we know what people want um so there are different algorithms for that I cannot say those words so I'm not going to attempt it but you can see them here on the right hand side we have dimensionality reduction this is where we're reducing the amount of data we retaining the data Integrity often used as a pre-processing stage and we have lots of algorithms for this principal component analysis linear discriminant analysis generalized discriminate analysis singular value decomposition uh Laden uh dial I can't

1:24:24 can't say that word there's just too many words that are too hard to say but there's a lot there's a lot for dimens dimensionality reduction um yeah and so hopefully you can remember those things classification regression clustering Association Dimension dimensionality reduction [Music] okay let's take a look here at neural networks and deep learning first defining what are neural networks so these are often described as mimicking the brain you have a neuron or node that represents an algorithm the data is inputed into the neuron and based on the output the data will be passed to one of the many connected neurals the connections between neurons is weighted the network is organized into layers there will be an input layer multiple hidden layers and an output layer you could technically have one hidden layer but often you have multiple layers if you have three or more now we're talking about deep learning if you have less than three that's just a neural network um and just look at the visual for here for a moment because each node or uh neural remember that it has its own um its own algorithm like how it's going to

1:25:30 process that data and I'm pretty certain that most neural networks the the algorithm is going to be same for all the nodes but we'll talk about that as we dig deeper into the neurons themselves um but then there's the concept of a feed forward neural network which is initializes fnn I don't know why it's not ffnn but whatever so these are neural networks where connections between nodes do not form a cycle that means that they always move forward so data moves forward okay we don't have neural Network's going back and this way and that way they're just going One Direction which is forward then you have back propagation this is where after um things ran into like everything's ran through it's going to move backwards through the neural network and adjust the weights okay to improve the outcome on the next iteration so after it's ran it actually just to update all the weights and that is back propagation this is how a neural network learns it has to do back propagation okay then we have a loss

1:26:38 function so it's a function that compares the ground truth to the prediction to determine the error rate so how bad the network performed ground truth right is data that is labeled that you know to be correct okay okay now we're talking about how these neurons are going to have their own algorithm right because up here we say that uh it represents an algorithm so this is where we have these um algorithms which we call activation functions so an activation function is an algorithm applied to a hidden layer node it's one of these things right here let me just get my pen out again one of these that affects the connected output and so an example of that would be ru or reu I don't know how to pronounce it properly but I recognize it uh but we will be looking at activation functions when we look at Neons a bit uh a bit soon here um there's the concept of DSE so when the network layer increases the amount of nodes we call it more dense uh and when the layers decrease the the amount of nodes we call it sparse okay so when we see an increase it's dense if it's

1:27:51 decreasing it's spse um and for deep learning algorithms we have supervised and unsupervised just like with classical machine learning um and so on the supervised side we're going to see things like uh fnn RNN CNN so you are passing in labeled data for this to work for unsupervised learning we we have uh dbn's SES RBM S's and not important to really remember this but I just wanted you to know that they have supervised and unsupervised learning um uh for for de keep learning [Music] okay hey this is angre brown and we are taking a look at Bert so Bert stands for bidirectional encoder representations from Transformers and this was a model that was released in 2018 by Google researchers um and you can find uh the models on hugging face for the account organization Google BT but the easy way to think of uh Bert or this is the easiest way that I like to understand it is that there's that Transformer architecture that we saw from all attention is all you need if you were to just take the encoder and then feed it into itself multiple times you would get

1:29:03 Bert and at least that's how I understand it to be um and so Bert is bir directional meaning that it can R text both left to right and right to left to understand the context of text so it really does understand um the context of of of words and where they are in sentences um Bert is pre-trained on the following task so it uses Mass language model um and so that kind of task is where you provide uh let's say a sentence and then you have like fill in the blanks so they say that they Mass tokens for uh sentence input but I would describe it as uh it's filling in the blanks for sentences the other part of it is it does next sentence prediction so this is provided this is where you provide two sentences um A and B and so the idea is that Bert will predict if B would follow a so by being trained on these two things um it provides it the ability to uh be aware contextually aware of sentences and be very good at natural language processing now the model itself in this state you're not going to do anything with it you need to then train it further for a specific task and so this could be for name entity Rec uh

1:30:14 recognition questioning answering sentence perir task summarization feature extraction beddings and more there's absolutely more things you can do here and Bert comes in many many different sizes so you have the base which is 100 million parameters you have large which is 240 parameters you have uh tiny which is 4 million parameters and there's like 24 or so other additional models um if you go to the Wikipedia it tells you how much data it was trained on it was something like 800 million something somethings uh and it was like the the Corpus text was from the Toronto library in the Wikipedia English so this is primarily trained on English I'm sure there are multilingual variants uh if we were to go over um to hugging face and uh there are many Bert variants that try to be better or solve other things uh with Bert um which is fine and while Bert is an Al model it is still used and it's ubiquit Baseline in natural language processing so when I say that I mean like they use it as a baseline or or uh against other other things because it's just uh so good um that it's not going away anytime soon here's an example of using us Bert so

1:31:24 let just get my pen tool out here so we can get a little bit closer here but the idea here is that here we're using um Transformers which is a hugging face library and we are loading in the sentiment analysis Pipeline and what's that going to do is that going to download the The Bert based uncase model that specifically trained for sentiment analysis now I said earlier that Bert it in its pre-trained state you wouldn't use it for anything right you need to train it further uh for those tasks and so it doesn't look like it but it is actually pulling a a a trained a very specifically trained one on this size of model and so here we can uh take these two sentences and then we can um uh do sentiment sentiment analysis on it and get a score back so hopefully um that makes sense there but yeah you're going to see Bert again and again and again so it's worth your time to [Music] learn all right so in this video I want to see we can get some hands-on experience with Bert so Bert is the bidirectional encoder uh representation from Transformers uh by the researchers at Google and so um I think that we can use this to create embeddings and that

1:32:35 is one use case for it um I don't know if Bert actually has anything to do with it but it's kind of fun uh seeing Bert there since this is a a Google researcher thing I think that maybe we might want to try to utilize either uh Google collab I wouldn't mind also popping into vertex Ai and seeing where we could run it there um but let's go ahead and try with Google collab first I am not experienced with uh bird at all but I'm sure we can stumble through it together and see if we can figure it out so I'm going to go ahead and make a new notebook and um this will be called uh I'll just call this Bert Bert uh embeddings simple and so all I want to do is figure out a simple way to do that now there is an option here to start generating code with Gemini so I'm going to go ahead and choose that um and say

1:33:23 you know I want to implement a simple embeddings example using uh BT okay let's see if it can do that for us I've never used the Gemini feature in here but that was really fast and so looking through this I've seen code uh around on the internet and this looks pretty much exactly like other tutorials um so like even off screen here I have one from like Chris McCormick and I think if I go down here you'll notice like you load it you have the you're loading it from uh hugging face that's where it's coming from okay and so similar things here so yeah this looks okay if you're wondering where this is coming from I'm going to I'm going to go um this is hugging face right so we're going you're going to see it a ton of times but I'll just show it again so I'm going to type in this Transformers hugging face and so this is pretty much what we're looking at right here when we're utilizing it so I think this is built into these are models that are easily accessible within hugging face that uh are so commonly use that they're just kind of plugged into here um but uh sorry I just was trying to grab the link and put in the slides for folks

1:34:33 later here I think what's interesting is they have a Bert Japanese so since I'm learning um uh Japanese language learning I'm kind of interested in this one and so this looks similar it might be a variant of it for embeddings you might try it for fun uh not that there's anything super exciting about it but let's go open this one and see if this is actually the example um that we have here because they might have a simple example here of utilizing it so here's BT base unase which is probably the most basic one that we can utilize no direct example oh down here we have one kind of but I think this might be fine so let's go ahead and take a look here so we have Bert tokenizer Bert model from hugging phase we're loading the pre-trained model called Bert uh base uncase and you know I keep saying that you anytime there's models on hugging face you can look them up right put them in here here and we can find them right so if we wanted to learn a bit more about this one we can read about it pre-train model on the English language using a mass uh language uh modeling MLM objective so here we can read about it and it looks like there are some variants of models that we could uh swap in here and then we have some code here as

1:35:44 well okay so nothing crazy I'm also just kind of curious for deployment um I was just trying to see if uh there was like a oneclick to Google collab but there's not that's totally fine let's go back over to here and so here we have tokenizer okay and then we're loading the models of the weight so loading pre-trained model tokenizer so the vocabulary then loading the the weights okay then we have our text that we want to um tokenize which we're basically uh yeah we're going to tokenize so we're using our tokenizer and then here compute token embeddings yeah I'm not exactly uh sure exactly what it is but um I'd say this is fine so we'll accept this I just want to go ahead and I want to see if this can changes so how can I print out well it has the sentence and Bings down here so it actually does it down below here so that's totally fine okay so I wasn't 100% certain but let's go ahead and see if we can run this okay so we'll just give that a moment and does it work well it's going

1:36:56 normally I'd break this up into separate lines but since we have this Big Blob here I just wanted to see uh what results we would get so I'll just wait for it to finish all right so it has completed and so we're getting back um all the embeddings you can see there is a a lot of um uh stuff in here so I would just be curious to see how we would run this on uh Google Cloud specifically invertex AI if they have a similar environment um I mean f obviously Google Cod lab is fine but I'm just trying to get it exposure to more toolings as much as we can um so I'm going to go back over to here actually maybe to the the B Japanese to me is kind of interesting so I might try to run this separately as well um so let's see if we can run that so I'm going to go ahead and just grab

1:37:40 this code here and by the way I'm just going to grab the link here put in notes so later on uh we can reference them but uh yeah let's go ahead and see if we can utilize this one so this one has Auto Model Auto tokenizer so we go back over here and it's a little bit different yeah it's a little bit different I'm not sure why it is but maybe because it's a more generic model it doesn't use Bert and uses the auto model I couldn't say for certain but I'm going to go ahead and add a new block here and we already have um torch should already be included above I think it is right if we go back to the top here I want to clear out this uh this box this one's annoying there we go um and we already have that imported so I'm going to go ahead and bring this one in and then we'll grab these ones okay we'll bring them down and then we will bring in our Japanese line here what does it say um let me see here that looks like Necco Des AI that looks like a Neco that looks like a cat

1:38:52 to me I'm not sure about the rest let's go find out that's the only symbol I recognize I am a cat Okay waga Haan NE that is that's a fancy way of saying it w okay I would just say I would say it a much simpler way but my my Japanese isn't that advanced we'll go ahead and grab the next line here and we'll go to our next line here and then we will grab its outputs examples of using the model character uh tokenization well hold on hold on hold on so this is an example of using a model with mecat and word piece tokenization okay so up to this point that's fine I don't think I care about the rest here let's go ahead and see if we can run this again I'm not saying I understand all the stuff I'm just saying like you can see how easy is to grab the code and get out values but again we need to really consult um an expert and so I

1:40:03 might pull in an expert here to have better context other than just starting to work with these things okay so we have a model error so do you install fugashi to use that [Music] okay so open this up in a new tab here and yeah we'll go ahead and just install this so I'll just bring this down a line here we'll just say um percentage pip install fugashi and I'm trying to remember like maob is um I think it's like a dictionary I used it before uh in preparing for the free geni boot camp but there were uh these things that if you gave them Japanese text it could break down its form uh based on a dictionary so I believe that this is a basis for that maob dictionary that it wants we don't see maob here it's somewhere underneath um as a library and saying this isn't installed either okay so we'll get that as well so I'll just dump it here on the end and eventually we'll get everything

1:41:11 installed now we might have to restart the run time sometimes you have to do that but we'll go ahead and try this okay but working with um anything with llms you're constantly having to fiddle with a library so it's totally not unusual for this to happen you could be writing this later on and you could be getting total errors and everything is um they're like Andrew why doesn't your code just work it's just the space is very um uh it's just how it is right so that Library looks like it is now installed let's right the next one and there we go so we got some output here and so it looks like we have um some special things or saying this is the start this is our separator so these are um specific indicators and then maybe I guess it's a identifying this as the section of the sentence um so again don't fully understand it but the point is we're able to get a run I just like to do one more thing here is if we can try to go do this on um Google Cloud so here in hugging face they'll have options like deploy and I rarely ever touch this I'm just kind of curious what would happen if we do deploy with Google Cloud code spaces now I have a Google Cloud account so you'd obviously have to make one and

1:42:19 we're in vertex Ai and it's opened up model Garden which I'm actually surprised that it's in here um so it says open models on hugging face now if you do not have uh if you do not have um any room for spend do not do this I'm just doing this to kind of explore it together so we can just together see what we can get as a result if you've never pressed that button so please don't do this because this might cost something and I want to have you to stay in the free tier so here it wants to deploy it to vertex AI which is the fully managed AI platform if you never heard of vertex AI you can think of it similar to Stage maker Studio Labs or Azure ml studio and then down below we're bringing the model we have our endpoint we have a place we want to deploy it there are the machine specs which there is nothing displayed here your endpoint will be deployed with the

1:43:03 following settings so it doesn't look like we get to choose but it's going to use a C3 High CPU for accelerate type unspecified um and then we' get an end point that we would hit but I'm not exactly sure how I would end up utilizing this um so I have to select something but I have no options here and maybe the reason I have no options is because I don't have uh access to this um this item here so I'm going to go ahead and just search for what this is in in Google like I want to know its cost right um so this one here is a C3 C3 Series so Zeon scalable processors fourth generation okay Titan titanium this doesn't seem like something that I I I would not not have access say too uh custom built security microcontrollers okay that's not gpus but it seems like maybe this is ried on CPUs and not gpus that's what we're always looking for like what are we running this on and so I'm pretty sure yeah this is running on this is running on CPUs the question is like why don't I have access to