Orientation Session 2

[00:00 - 00:50] Just a little bit of introductions, basically you guys know my background and it's the idea of the course is to basically for to build to be able to network, to be able to meet people and so forth to kind of buy every lecture. You can go to the afternoon one or the evening one, depending on what you prefer. Why don't you guys provide a little bit of intro on yourself, which industry you're from and just as a kind of brief intro who wants to go. Yeah, I'll start. I'm James O'Connor. I'm a front-end app patient developer at SAP, working in the travel domain, mostly, or specifically. And on the side, I also choose scuba. I mean, that's of other interest. I've built a little bit on my own with just five coding, really, an experiment a little bit with Langchain and connecting it up to a model, but not anything major. Excited to get started. All right. Thank you, James. Who else wants to go?

[00:51 - 01:32] I can go next. Yep. I can go to. Hi, I'm Aditya. I've been in the industry for a couple of years. Most recently, I was with Google and I work. I started out with infrastructure software. In the last couple of years, I did spend some time working on a few L.U. models and also with the end, after the advent of gender, the way I also played with some of the models that might work. I currently am with Walmart, but that's my background. Hi, Gonzalo. After companies, I'm mostly on the IoT industry and short-term rentals and residential.

[01:33 - 01:47] Yeah, looking for ways to bring AI into the mix. I've been an AI researcher, say, 11, 12 years ago, when none of this was even possible because of the computing limitations. Yeah, I was hoping we could do what we can do today back then.

[01:48 - 02:51] Yeah, go ahead, Julie. Yeah. Hi, my name is Shulir. I'm a software developer working mostly left for only 10 years on the wrong time. I'm in Toronto, Canada, and mostly working in banks and venture plans. But I'm really curious always about technology and consulting on young and my son who is 18 years old. She is in the AI for a very long time and not just plan, but she is building stuff. So she knew. And I was looking at my entire site and sometimes so far behind him that I also want to learn and maybe somehow figure out how to use my strong tech expertise to get right and build your application. Thank you, Julia.

[02:52 - 03:53] Hey there, my name is Andrew. I live in New York City area. Currently, a full- stack engineer worked in on in the healthcare industry. Currently, I really don't use AI and machine learning that much to fitly co-pilot with my editor and just taking this to get jump-started in that learning process. All right, thank you, Andrew. And yes, I'll go. I can't tell how many people there are remaining. I'm Craig from the west coast here, and I've been in tech for 20 some years. Majority of the latter half has been mostly front-end things, but about 10 years ago, did work at a small company where we were doing machine learning and getting back to that and to see where the modern tech has taken us . So anyway, thanks.

[03:54 - 04:32] All right. Thank you guys. So the purpose of the the coronavirus lecture is orientation. So this course is unique in kind of about two aspects. One is we have a course component, and then we have a coaching component. The idea of the coaching component is to be able to allow you to be able to complete your project, whatever it is. And we're going to go into the specifics of the projects. But so this is basically providing an introduction for everything . And the goal for today's session is to address your concerns. Feel free to ask questions as we go.

[04:33 - 06:24] The first thing is basically understanding how it's structured, what resources are available, how to engage with peer-to-peer, as well as the community, and then how to be successful with the course to get what you want. Please, questions as we go, I'll stop to address questions once in a while. And then at the very end, we'll also have FAQs as well. So the idea is to introduce everyone, provide an introduction to the course, introduction to the community, introduction to resources, a little bit about learning what to learn. A lot of the feedback that we've done has been that this course is very intensive. We're going to basically go into exactly how to basically navigate it and deal with it in a way basically that fits with within your kind of work lives. Some of you guys, they will like the intensity and that's actually what you're looking for. Some of you will basically have more, I don't know, family obligations, and we'll have to basically be able to fit it in probably. Then we'll go into a little bit of the deep work and distractions and tools that you can use as well. I mentioned kind of about my background before. So part of the start of this is we started creating LLM agents internally. We used it for reaching out to people, we used it for sales tools, rubric tools , editing tools, and other things. So as we basically went through it and started building agents, we realized that the material outside in getting to the state of the art, kind of about agents wasn't very good. And it was very hard to have a cohesive kind of a place where you understand the foundational models from basics to the state of the art, as well as understanding agents from basic to state of the art. Dippinen is here as well. Basically, he's an AI kind of by ML researcher and then he helped create the lectures as well as the code as well.

[06:25 - 07:03] We have one of the things that people really like about the lectures is there's a lot of examples, code examples that are relatively up to date. So this is Alvin. So I'm sure you guys have seen the fundamentals of transformers. And we were working on a course with him while he was at Apple. But and it wasn't due to his kind of schedule. It wasn't quite verging properly. So we basically changed it to a workshop and in the workshop basically inspired the book camp as well. So he's available for Q&A as well as yeah, permanently kind of a Q&A and be able to clarify conceptual questions.

[07:04 - 07:34] So Marianne, I'm sure you've gotten information from her already. She's a lot of the community manager, operations manager, assistants from kind of out different things. You can either email her or you'll you can message her on our community website as well. In the last cohort, these are the set of things that people have built. Some of you guys have been on the webinar basically where I mentioned this. Some of you guys were in the initial webinar when I didn't mention this. A number of these things are they're all agents, but they have different techniques behind it.

[07:35 - 07:51] Some are fine-tuned, kind of a model. Some are multi-modal fine-tuning. Some are text fine-tuning. Some are code specific, basically building landing pages for little businesses. Some are multi-modal fine-tuning for detecting but things on Facebook marketplace.

[07:52 - 08:29] One thing is a rag, kind of a chat with servants for churches, document processing for insurance claims. So part of the reason why we're basically emphasizing the project is no matter what you guys want to basically do. Generally speaking, there's three types of people basically here. People that want to do something for work and advanced career. Some people that want to do a startup and other people that want to build the AI consultancy. So no matter what it is, basically this will help you consolidate a lot of the information and then be able to demonstrate kind of a career capital. An example is basically the very last one .

[08:30 - 09:00] The very last example is that someone built a text to Qatar tab generation. And this guy was basically a guitarist on the side. He was the hobbyist and then he basically kind of built this. And now he's at Capital One generative AI, kind of about group, basically doing generative AI, basically. So he basically credits the course as well as kind of just being able to use his project and be able to utilize a lot of the notebooks to be able to create his project.

[09:01 - 09:22] So a number of these things are legal aid assistant that's basically someone from the US government, personal job search side developer commercial real estate assessment using AI. Basically, that someone was hobby was a commercial real estate. In three and a half months , you should be able to build a agent. Basically, an agent is a broader term now, basically.

[09:23 - 09:42] And there are specific techniques for agents and patterns for agents. But in general, it's like saying you're building a web application. So an agent basically uses tools. It can resume about things. But in the underlying kind of a layer, you still have to use prompts. You still have to use vector databases, which you will augment a generation.

[09:43 - 10:42] And then you still have to use fine tuning. The second thing you should be able to do at the end is be able to evaluate. So a lot of developers take unit tests for granted, like when you're basically prototyping, you're basically about coding during the rush. And they basically take the same approach for evaluation. And whereas evaluation is much more crucial in AI applications. The reason why is there's a lot of drift that happens, basically, as you build agents for long-term tasks. There was a guy basically that building an accounting agent for over the last, I don't know, six months or so. He basically published it on Hacker News. And what he basically noticed is that the data as well as the application started drifting and it became more and more inaccurate over time. And so if you don't basically have evaluation, the surface area is natural language. And you're just unable to know whether you're getting the same experience as your users. And then, of course, the cost as well as the ROI.

[10:43 - 14:12] The reason why we emphasize the project is what hiring managers and customers basically buy from you is a combination of skill depth and portfolio proof. Being able to prove that you can do something with working artifacts as well as metrics and about behind it. There 's some people basically that said, okay, when I basically mentioned the webinar and they had clients, they said, yeah, a lot of clients basically say to me, all I have to do is give me this customer service log, sales log, and I basically have to improve it somehow and improve it. And so that is the portfolio proof creating a working project. And then the second thing is it's linked, obviously, that's the skills as well. The distribution and people that care about your proof is really dependent on what your overall goals are. Basically, some people have in the previous cohort, they've they got promoted within their company. So for example, there's a guy what I caught up with last week. He basically is now leading us a three person team that's building RAG and or customer service. And so you can basically distribute it internally, you can distribute it to customers, you can distribute it to existing clients, local businesses, and so forth. And the story is being able to crisply articulate everything. So the distribution and story is basically highly dependent on the individual. And but what we're basically focused on the project is so the project is your portfolio world that demonstrates your skill tab. I want to basically mention that AI engineering is different than foundational model engineering, though we will cover a lot of the concepts that are inside foundational model engineering as well. foundational model engineering really basically takes a transformer based LLM, and then uses reinforcement learning to fine tune it and adapt everything. AI engineering is basically taking the same kind of model. And then you have the capability of slicing things from the model, putting lenses on top of the model , and being able to adapt it with your own data or dealing with it. So the process of fine tuning is being able to perform surgery on the model. So you can basically perform surgery and cutting things out from it, basically, or putting things on top of it, or you basically fine tune it or your instructions, your concepts, and other things. And then enterprise AI engineer is not unlike AI engineering, but you have more security kind of concerns. So I'm sure you guys have all dealt with security kind of issues. But the issue with natural language is it's much wider in terms of security in terms of so we're going to go into like how to hack LLM, basically, because the reality is basically a lot of fine tuning a model or dealing with a model basically on a security basis is making sure both on the interaction kind of paradigm, you have sufficient guard rails around certain things. People will basically try to hack, make your AI robot behave differently. They'll basically say crude and delude things to it, basically, and they'll try basically once or twice in different different directions. And you can actually see this in anthropics, RLHF, basically, instructions, basically, you can actually go through it. And then there's a lot of that from force that are like very crude and lewd and profane basically in it. And so the reason why it basically is trying to teach the model to basically avoid those type of conversations.

[14:13 - 14:38] And so that's important basically for the enterprise and also for kind of a normal AI engineering. And finally, basically, a number of you guys are interested in AI consultancy as well, being able to demonstrate your case studies and being able to know which tools to apply in which case is going to be informed. The final thing is that AI startups. You may or may not know, basically, but a lot of the AI startups basically use AI engineering stack.

[14:39 - 16:17] From the surface level is looks like a bunch of people are just using the opening API, but they're actually using, for example, when we talk about Windsor or or cursor, basically, they're using a chunked kind of a version based on the abstract syntax tree, and they're using a graph database, then they're using a fine tuning embedding model, basically. So it's an entire stack of things. And they're often fine tuning open source L and as well. So one of the things that we basically kind of have in the course, and Marianne will basically get your preferences is accountability partners. The accountability partners is the idea is basically to be able to have someone to be able to check in with, have a conversation. And this is not just basically saying like, Oh, do you do your homework or not? Basically, people basically say, Hey, I'm getting confused with this thing. Basically, I tried this model for my PDF. It's not working properly. And the idea is to be able to manage you guys basically to different people. And so in the last cohort, the number of accountability partners, some people basically form companies from it, some people form consultancies from it. Basically, there's a number of things that can basically happen as a result of this. You guys are adults, so we're not going to force you to do anything you don't want basically. So we're going to ask you for opt in and for no preferences, basically, and then we'll match you as needed. And finally, the community, I've done a lot of courses and projects, a lot of courses and projects, they don't really have a community component, basically. And so that so what we basically created is many projects that are done with groups. And these many projects are to simulate real life, basically. So what is real life?

[16:18 - 18:26] You get a set of customer service logs or sales logs, and then about the boss or the client or whoever basically says, Hey, go build something, basically improve this, basically. So we're going to have something like a cooperative competition, basically. So where people are competing to basically get the best evaluation score. And why are we basically about doing this is you 'll basically kind of see that data science and machine learning is about utilizing the different kind of techniques. And then, for example, we basically produce 20 rag techniques. And then, for example, team one basically uses one, two, and five team two uses 13, 14, and 15, basically, though, produce kind of by different results. And then at the very end, we'll basically be able to show kind of about the person that had the team that basically had the best results and their notebook to be able to sell other people can basically learn from it. Then basically, there's a Miami event, it'll be on the weekend, and you'll be able to network with people, share techniques, basically share projects that you're basically doing, collaborate on projects you're doing, and so forth. Finally, basically, the coaching aspects. We've had the onboard ing about call with you guys. And in particular, one thing we're basically changing about this cohort is we're gonna have more one on one calls with you guys to basically get your project on track. The problem with basically the previous cohort was that some people weren't able to know which projects, basically, they wanted to do until they understood the techniques. And then by the time they got to the techniques, it was already too late, basically. And then some people basically just got stuck in ideal mode and other things. We're gonna basically have a document that suggests we'll document your progress and then suggest videos, basically, for you to take a look at and to progress on different kind of things that you want to do. And then please also tell us kind of about your kind of different preferences. So you may or may not realize that as we go through the course, but a lot of our projects are very different than traditional academic machine learning projects.

[18:27 - 19:12] Basically, part of the reason why is we adapted it to real life projects that people have done. Basically, in the previous course, we had three insurance people manufacturing company, a two banking kind of guys, multiple startups, portable tech business owners. And they effectively basically said, we basically need this multimodal thing. We need a better LLM- based web scraping or other things. As you basically go through the course, tell us what you basically will you basically prefer. And if you basically prefer us to put research papers and our project techniques and other things, and then we'll try to accommodate it as much as possible. And then we're gonna basically initiate the other thing we're basically doing in this cohort is the initial project coaching.

[19:13 - 19:53] So I'm sure you guys have seen the lectures as well as the project coaching invitations. The initial set of project coaching will be lectures. So one of the things that basically happened in the previous cohort was that people didn't have a sense of the proper different types of AI startups that are basically happening. So AI startups are basically very different than. So we've gone through three generations effectively, basically all four goal of the mainstream stuff. But like we went from desktop, kind of about PC to web, web slash mobile, and then we're basically transitioning to AI. And whereas AI, basically, the most the interaction paradigms is basically a bit different.

[19:54 - 20:40] So if you think about basically what happened with the internet is every single person has 15 to 30 to 50 databases that they're working with on a daily basis. And you might be wondering, I don't work with a database. If you're it, if you're using zoom, if you're using notion, a lot of the databases that basically abstracted away from you. So with AI, basically, there's another level of intelligence is the intelligence on top of the data. So the idea is basically that we're going to have digital clone digital intelligence, digital coaches that are associated with basically every single person and to augment people's capabilities, but also to basically as interface basically for people as well.

[20:41 - 21:03] That's just an idea of basically talking about the project types and project templates. That's just one project template, basically. So we're going to go over a lot of different examples of AI startups that you can basically build like citation engines or kind of search engines or and most people think it's just, oh, I'm going to basically create a chat bot and attach it to whatever domain I'm basically kind of by doing. And there's much more than that .

[21:04 - 21:21] Part of what we basically kind of build in the book is being able to create material to be able to work within your working memory. So your working memory is like your computer RAM. Basically, everyone's about computer RAMs are larger these days. It's 32 gigs, 64 gigs.

[21:22 - 22:47] And your working memory is seven. You know, basically, not seven gigs, not seven megabytes, but seven, basically. So the reason why traditional academic education is so long is it basically your bandwidth on concepts is basically limited by your working memory and your ability to be able to utilize kind of long-term memory and have things basically turn into intuition and be able to manipulate concepts is your ability to be able to manage the bandwidth and encode it into your long-term memory. So one of the things that we basically do is we manage your cognitive load, basically, so as much as possible. And so there's multiple kind of about cognitive loads, but we basically tried to maintain your cognitive load. So we basically put analogies that are relevant to your day-to-day or software engineering context. So we'll explain a lot of things in everyday terms primarily so that you can anchor it already and be able to utilize it basically natively as a first-order kind of object. But the only other thing I want to basically mention is despite us doing the best, you guys will feel overwhelmed, basically , with the concepts. Basically, that's the universal thing that we've heard from a lot of the people . Basically, so you will have to be able to manage this cognitive load, use ChatGP, use a system prompt.

[22:48 - 23:13] Basically, in ChatGP, there's like a personalized bus section. There's what's known as a system prompt. Plus, put in there, basically, explain it to me via analogy in whatever context that you basically want. And managing the cognitive load is going to be fairly crucial for this course. We basically provide about not just lectures, but coaching. And in particular, we want to talk to you one-on-one every two weeks for 30 minutes about your project.

[23:14 - 24:12] Be able to, as you basically go through the project coaching lectures, you'll be able to understand the different types of AI startups. And just like how the internet transition to mobile, people basically started saying, oh, I'm building, people started understanding on demand. So I'm building Uber, but for groceries. I'm building Uber, but for, I don't know, local alcohol delivery or whatever it be. And that basically became an analogy based way of thinking. Basically, it's both from a first principle standpoint, but also from analogy, you'll be able to say, I'm building a digital clone for real estate agents to facilitate 24/7 kind of interaction and transaction, basically. And yeah, so basically, we basically have lectures at the very bottom. Then we have exercises to facilitate drilling. We have the community coaching and we have the project as well. So yeah, so this is basically going to be, we'll show you the exercises in a little bit so you can take a look.

[24:13 - 25:05] So what is AI engineering? AI engineering is basically the evolution from a long time kind of algorithms back in the 1970s. Prior to the, prior to the current large language models, a lot of the things that you're basically doing with data engineering is you have to deal with a lot of data processing, a lot of feature engineering and attribute engineering. And typically, you wouldn't be able to do it unless you're basically at a big company, simply because it's very hard to collect all the data and be able to create the type of ML application that you basically want. So what we are right now is we're at the react moment of AI. So we react, so if you know anything about the web programming days back in the day, it was HTML, CSS, JavaScript, it was kind of out the well west, not a lot of standards.

[25:06 - 25:48] And then react came along and basically had virtual doms. It basically had a prescribed way of doing things. You can basically download things from the internet from other people. So part of the reason why we're able to do teach AI, basically without a lot of the math, the calculus and the linear algebra and everything is a lot of these things are abstracted away for you in libraries. So we will teach you basically the concept behind it so that you understand these things. It's not that you can get away with not knowing math entirely, but a lot of these things are basically abstracted away, just like how you don't deal with spread black trees or trees or distributed hashes basically day to day, unless you're a lower level engineer.

[25:49 - 26:32] Basically, this is basically fun about the key thing from dealing with large language models. So in this course, we're basically dealing with primarily transformer based multimodal large language models. The reason why I basically mentioned that is that when people basically talk about generative AI, there's generally two types of, there's three types of generative AI, but there's primarily two types. Basically, you'll basically hear people talk about diffusion based models or transformer based models. Basically, diffusion based is a different technology, basically. And so you basically ever use like mid journey, basically to create things where you basically see it, where it looks like a blurry image and it gradually kind of fades into talking about becoming clear and clear. That's basically called a diffusion based technology.

[26:33 - 27:49] What we're basically learning here is transformer based multimodal systems, large language models. We're not learning GANs. We're not using diffusion, basically, but it's transformer based large language models. So the reason why is you can't really build agents with the diffusion models and transformer based language basically infuses language natively in a way that allows you to use tools, be able to allow you to reason things and be able to do things with them . And the other thing that basically that transformer based multimodal systems is it has a be built. You can call it like software sensors. So it has the prebuilt kind of capabilities to hear to see, basically, to kind of speak as well. So most of that was basically kind of the foundations of a lot of that was done in 2016 with deep neural networks, basically. And now we're able to basically infuse everything with language. We basically talked about kind of about the evolution , but what we're basically kind of about doing is everything is going to be infused with prompts and languages, basically, and these language and these prompts basically kind of about control , both the workflow, it controls the modality, the image, the video, or whatever.

[27:50 - 30:04] And it controls, it's able to reason about things. So what we're basically fundamentally doing is we're basically using prompts to interact with data. And then we're basically calling APIs. And so this class of software, basically, that's emerging. People are tentatively calling it vibed software. You're vibing with it. Basically, you're basically associating. And the reason why people basically say it's a vibe is it's probabilistic. It's not deterministic in the way software is basically. So if you write a prompt, basically, the exact same prompt, there's actually a randomness, basically, even in chat GPT. And the early kind of vibing is basically vibed coding, basically cursor, augment, kind of a lot of these companies. And then it's starting to shift into other areas as well, vibed marketing, vibesales, and so forth. What's really important to be able to understand the concepts in the foundational model, and be able to understand the libraries, and then be able to use the techniques, be able to utilize kind of about everything. So we talked about basically the react moment being here for AI engineers. So we are learning some of the things basically behind the foundational model engineers, but not everything, basically. So if you want to go to foundational model engineer, you can with a lot of this training. But fundamentally, what we're focused on is AI engineering, basically, and we start with evaluation based prompting, we go to synthetic data. We then go into how to build a search index. So you may basically think, what do you mean a search index? I thought it was for a tree will augmented generation. A search index is basically being able to use traditional databases with vector databases, AI databases, and be able to use a large language model. This is basically known as retrieval augmented generation, but people basically make the mistake of always assuming that it's a vector database with a large language model. The reality is you can attach a lot of different things to it. You can attach graph databases, you can attach a lot of different things. And then we're going to basically go into agents and we're going to go into fine tuning. So most people basically think fine tuning is just a one level fine tuning. We're going to go to I think is five different different techniques for fine tuning.

[30:05 - 31:51] And then at the very end, you'll be able to build a live application, be able to understand your evaluation, like true positive rates with precision and recall, and and be able to have a story behind it. You're also going to basically enforce your muscle memory as well as the concepts, basic rules, homework activities, places, your own project. It's very important that you be able to manage your load basically. And this in the last cohort, basically, there's two ways that basically kind of students were able to basically do it. One is people manage pick their battles basically on which ones they were able to focus on and to do some people basically did everything. Basically, if you do everything, it will basically require probably 25 hours per week basically to be able to do everything. Some people actually did it with with a family. So the way they were able to do it is they basically just woke up like at five a.m. in the morning, basically to be able to do it. But I think in the last cohort, like there was like three people that did that basically. And don't be guilty if you're if you're not finishing everything, but you really want to be able to understand basically the concepts basically by looking at the code and ideally going through some of the exercises. Before we go into the resources, did anyone kind of have any questions so far? A question there when you said you can attach databases, is it possible also to attach like generic relational databases ordinary? Yeah, yeah. So basically, what you basically see is that so retrieval augmented generation is the merge of retrieval, which is search and generation, which is large language models. And once you basically get into production level techniques, one of the key things that people basically do is they use a metadata filter.

[31:52 - 34:28] Basically, so you're filtering on dates. You're filtering it on a lot of traditional relational database things. And so you actually see the convergence of databases in that relational databases are adding vector capabilities. And then then traditional vector kind of databases are adding relational kind of types is because you basically need these like filters. And so at a very high level, basically, so you have this retrieval augmented generation. And you can basically have relational data filters. You can have a graph database basically. And so these were tree walk the generation can actually are like compound systems where you add five or six kind of about things basically dealing with your private data. Because the reason why I mentioned that is the typical conception of RAG is basically you just add two things, which is generation and vector database. And that's like 2023, about the naive RAG that you'll basically see by Googling some language and tutorials, the modern kind of production systems that have evolved beyond it. All right, this is basically resource one. You should have access to this. If you don't basically email me or Marianne basically to get access to this. So this is basically a set of databases. And you might be wondering why do we need kind of open source models if I already know all the open source models, these are open source models. And they're also I tuned basically models as well. Basically, there's different types of capabilities that you can basically have here. Let me zoom in. You basically have different types of chat, multimodal image, audio, video, modality has to be put inside the model. And it 's very explicitly trained, basically, and then other capabilities, like additional data, whether you're talking about cybersecurity or software development has to basically find you. And so there's a number of people that have already done this on the web, basically, and have shared their kind of bio data sets. Sorry, they're fine to models. And then the second kind of aspect is data set. Basically, this data set is going to basically be important. For example, we have different data sets across software development, cover letter, data set, resume data set, UI reasoning data sets, reasoning data set is basically where you show intermediate examples . Basically, it's effectively, if you remember, when you basically learned a for loop, basically, how you learned a for loop is you have to unroll it. Basically, so a reasoning data set, we're going to get it into a later cut. And then come about their software development.

[34:29 - 35:27] Basically, there's Python co-generation data sets, Q&A data sets, basically, co-classification. We have cybersecurity, intrusion, anomaly detection, fraud detection, algorithmic training AI on historical financial data, clinical prediction, patient outcomes, clinical research, medical reasoning data set, traffic flow analysis, and equipment failure, real estate, market trend, housing prediction data, customer chatbots, and so forth. In this course, basically, we really emphasize synthetic data. So, synthetic data is not only basically going to be important on your evaluation side, basically get evaluation up and running, but it's also to basically augment your data set. So, a lot of times, you won't be able to find the data set, basically, that for your specific application. So, you'll have to scrape it from the internet and/or augment it with the synthetic data. And then, and then finally, is basically, and then finally, second is cloud services.

[35:28 - 35:49] So, most people basically think, oh, we have AWS, what we have this, why do we need cloud services? Cloud services are for the different parts of the AI lifecycle, basically, and there's cloud services around fine-tuning, basically cloud services, where they have server less GPUs, and they spin it up on demand. Basically, there's cloud services on inference.

[35:50 - 36:41] There's cloud services for reinforcement learning as a service. Basically, there's cloud services for decentralized training. Why decentralized GPU training? Because it's cheaper than AWS. Basically, there's things for data labeling, model evaluation, basically, since that of data generation, AI webspaper. Why AI webspaper? It's basically because I'm sure you guys have all built scrapers. Before, it's a pain to basically maintain. There's a new generation of LLM-based web scrapers that are less of a pain. There's AI capture solving, basically, there's GPU hosting, and then AI modeling and running. Some of these inference companies, they actually have their own proprietary summon conductor, and therefore, their inference costs are basically much less. So, there's actually different reasons to basically use cloud services, basically, these others in AWS and the big three, right, Google Microsoft and Amazon.

[36:42 - 37:29] Then, there's open-source frameworks, basically. So, just what we talk about the data set, there's a set of emerging, open-source frameworks, agent frameworks, autonomous agents, prototyping, deployment, data processing, benchmarks and evaluation, orchestration, fine-tun ing, retrieval, as well. So, different ways to basically combine two different things, basically, on the open-source and about the side. Then, here's the advanced, combined AI concept, different resources, basically, that you can utilize, basically, we're going to be adding more to this and classifying this. Right now, it's a little unorganized. So, this is basically about project ideas, and you can basically see machine learning, YC, basically, you can see YC generative AI. We're going to be adding more and more to this.

[37:30 - 37:47] We have indie hacker, we have starter stories, and then, where the idea is to be able to get inspired, basically, from what other people are basically doing. Then, we basically have newsletters and channels, and then, kind of, different newsletters, different research, different YouTube, different Reddit, different, kind of, self-stack, as well.

[37:48 - 38:49] And then, this is, basically, automations for an innate end. Some of you last, basically, are focused on building AI consultancy, basically. So, the issue with, basically, AI consultancy is that you can stitch together the workflow, but the problem with the workflow is you still need AI techniques in the back end, basically. So, this will become more valuable, basically, as you learn the different techniques, and then, you're able to stitch together these things. These are sales techniques. There are sales automation. It's email automation, DIN, Instagram automation, YouTube automation, and so forth. So, that's, basically, kind of, about the resources of this section. So, this is the community website. You might be wondering why we basically built our own community website. We originally used some other about community website, but the problem that we basically got was AI, which is not natively integrated into the site, basically. So, it was hard to find things, basically.

[38:50 - 39:00] We're going to basically build three, kind of, about core features into it. One is video chat. Second is digital clone, basically. And third is tools to basically reduce information overload.

[39:01 - 39:59] Basically, I'm sure you guys are part of Slack and Discord, where it's just a ton of voice , basically. And in the last call, where we generated up to 100 hours of content for coaching community, happy hours, and other things. And sometimes, we had valuable information in some of these hours, basically, but not everyone can basically attend all the calls, basically. So, the video chat is just getting designed to facilitate that. And then, we're going to have a digital clone of Alvin, a digital clone of DIPN and me. That's going to be trained on our transcript and our writings, basically. So, you'll be able to ask it, kind of, about questions, in case you don't want to, basically, in person. So, this is basically, kind of, about the onboarding. We want you to basically, kind of, go through the onboarding, be able to complete your profile , be able to save below, basically, add the lectures to your calendar, Q&A section, the coaching section, be able to access the recordings, click this to be able to see what the details are.

[40:00 - 42:58] Here is, basically, where you want to, basically, save below, introduce yourself a little bit about your background, your superpower, basically, how you can help as well. Here is announcements, basically, where we'll have updates, news, program highlights, and so, of course, basically, I think it goes without saying, please be respectful to other people , basically, I'm sure that's not going to be a problem, but, and this is general questions, just in case you have general questions, any questions about orientation, basically, and then we're going to have very specific Q&As, basically, for lectures, as well as for the coaching call, and happy hours, and then for the guest lectures, as well. So, that's, basically, kind of about the, the high level idea, then this is, basically, kind of about the courses section, basically. So, we have 18, kind of, mini tutorials, basically, for you , on the basics of getting started with, kind of, about platform, Python, introduction, basically, installing Python and virtual environments, basically, using Jupyter notebooks, and codes and exercises, we also, basically, introduce you to a hugging face, basically, different hardware for AI, basically, advanced AI concepts to be able to use, and then we, basically, introduce you to prompting and using chat GPT as a research assistance. There, there's three AI tools that you'll, basically, need to use. One is, I thought I didn't need to, basically, say this, but there were actually people that didn't subscribe to chat GPT Pro, basically , before, basically, but chat GPT Pro and/or Gemini, whatever you basically prefer, has a deep research functionality. Basically, it's, the deep research functionality is useful when you basically are searching for datasets, is, it can basically find, go through the internet, and it uses something called multi-hub reinforcement learning technology, basically, to be able to do things. The second thing is, basically, you'll need to, basically, use the system prompt, basically, inside your chat GPT, or Gemini, modify your system prompt, to basically be tailored toward you, and then really utilize it as a brainstorm ing assistant, and really become much more proficient, basically. So, we, we try to, these are not just, like, basic tutorials. We try to utilize all the previous kind of cohort lessons, basically, and a lot of the things that people had problems with, or our pro tips, basically, into these lessons, basically, and so these are 18 lessons, basically, for onboarding. Then, we're going to basically have different modules, basically, that we basically kind of about mentioned, basically, introductory module, kind of about unit 1, unit 2, unit 3, unit 4, unit 5, and so forth. We're going to have events here, basically, all the calendars are going to buy events, and then we're going to have members as well here. So, we're going to invite you, as soon we were basically rushing to kind of finish this, basically, or not finish this, it was already finished, but we were trying to basically spot all the bugs, basically.

[42:59 - 46:32] So, you guys know how it is the last 20%, 80% of the work, and basically, that 's the general idea of the community side, basically. And then, this is, we basically have three types of notebooks. So, this is on collab.research.google, basically, like, basically, what you want to do is you want to be able to utilize things on collab.research, basically, and there's three types of notebooks. We basically have exercise, we have notebook examples, and then we have many projects. Basically, notebook examples are where, basically, there's a code that you can basically run through, and then you can just follow it. Exercises are really designed for specific kind of things. So, I'll give you an example. So, for example, this one is, this is basically, kind of a unit 1.4, basically, a couple of lectures head, basically, and it shows Facebook OBT 125 million model 1.3 billion model, and it gives you a note on RAM versus VRAM and why usage will basically grow, how you basically do memory, kind of, clean up, and preventing memory buildup, basically, and then it shows you basically how to deal with it, and then here is basically, introduction to LM inference, basically, the basic pipeline, be able to prompt, tokenize, model, and decode, and then we basically have the exercise here. So, we basically try to make this exercise relatively fun-ish, basically, as much as AI can be fun, basically, so, you know, this welcome to an AI oracle room, where a powerful language model is here, build a complete LLM pipeline that can question process and generate a thoughtful answer and be able to print it out, and then you're going to basically utilize the libraries, and so, one of the things that people liked about our notebooks was basically like we have an over-abundance of commons, basically, and we try to introduce you basically to the libraries as well as to the individual functions that you would want to basically utilize, basically, and then, kind of, this is step one, this is step two, this is steps three, and so forth. And then , topic two, exploring model variety, basically, and then we have the exercise around here, where you basically say, who tells this for you the best, and then you basically both rule different examples here, and then, like, for example, storytelling, oh, here's what I wanted to show you guys. So even with something like, like, for example, tuning model parameters, basically, model parameters are like temperature, top K, top P, max new tokens, but since we first introduce you to low temperature versus high temperature versus the K most likely, next tokens, and then we basically give you the actual standby exercise, basically, and then we make it, like, fun -ish, basically, so this is a focus chef, this is a creative chef, minimal chef, balanced chef, and we basically then introduce you to basically a full-working recipe generator, basically. So that 's basically the exercises. This is new for this cohort, basically, because we went through, we actually had a lot of this, these concepts, basically, in the previous cohort, but we realized that giving people just, like, top K temperature and all these things, like, people didn't really internalize it and really basically, like, practice exercises sufficiently, basically, yeah.

[46:33 - 49:01] So that's basically about it. Oh, Craig, basically, you have to be invited from us, basically, we're going to invite you, we'll invite you very soon, basically, either later today or tomorrow, basically, into the community. The purpose of the course is to be able to do the following, basically, so most people, when they think of school, they're basically thinking about regurgitating things that the professor basically did. That's the lowest level of understanding. Basically, certainly, what we want to do is, basically, we want you to have an intuition and be able to memorize the concepts, but really, what we want you to basically be able to do is being able to move up the blooming taxonomy, to be able to apply it, to be able to analyze it, compare different techniques, basically, evaluate it, basically, and finally, create, basically, so the idea is to basically move you up the entire stack, basically, and this is the purpose of the course. Make sure, basically, we provide analogies, basically, in the course, a lot of, there's going to be a lot of concepts, basically, so as you basically go through it, build the analogies into your system prompt, so it just naturally explains it to you. Basically, some people have mentioned that, during the lectures, they have chatubd on, basically, and they have the transcript on, and then they just copy and paste about different things on the lecture notes and other things. We have to clarify it. The other thing that other people have done, basically, is some people basically download the lectures ahead of time to basically read the lectures, and so they basically, when they're actually inside the lectures, basically, they can ask more intelligent questions about what they do or don't understand. So, obviously, this depends on your time, basically, and so forth. I mentioned, basically, about the working memory using analogies, chatubd, basically, reading it ahead of time. Really, you really want to be able to manage your conceptual overload, basically, knowing what to focus on, other foundational model, adaptation, and other projects is also important, and then being able to maintain a deep work schedule. So, a deep work schedule is some type of schedule where you're able to work on and interrupted, basically, and then, of course, you need to basically be able to manage your project scope as well, basically, that's for any kind of project. Please make sure you really block off kind of about time or different things. So, there's two ways to basically talk about consult with the lectures.

[49:02 - 50:07] Some people, basically, obviously, attend kind of about live. As we basically go through the lecture, some people get off a word, they have to listen to the recording. You can also listen to the recordings, like a podcast, you know, basically, and then, so whether you're walking somewhere, or whether you're doing something cardio, basically, you can really listen to a lot of the things, and then be able to schedule time where you can actually sit down and actually do the exercises, or whatever your projects are. So, you may think that this is basically kind of obvious, but, and we actually mentioned this in the very beginning, but if you don't do this very explicitly, you're going to get overwhelmed, and you're going to fall behind, basically. And so, most generative AI kind of things are like a master's program, whereas, basically, two years, basically, we basically kind of condense this for three and a half months, basically. So, it's cognizant of your time, but the side effect or the trade-off of that is you do have to manage how you basically deal with this, basically, otherwise, you'll fall behind. And then, you want to basically become much more of a pro user of large language models.

[50:08 - 51:03] So, I think it goes, I just assumed this would be true, but you really want to basically do the system problems, and really be able to adjust, basically, use deep research, use research tools, use LLM-based tools. You may or may not have noticed, but I'm using an AI browser, basically, it's called a browser called DIA. Basically, there's another browser called Comet, basically, and there's a couple of agent browsers, but you really want to basically utilize it, not just basically as a gimmick, because they really improve your ability to summarize information, get through information much more quickly, and be able to search for a needle in the haystack. So, for example, the text guitar tab guy had a problem planning his dataset, and he basically did traditional Google search, and his breakthrough was, basically, when I suggested using Gemini's deep research, and he was able to find the exact dataset that he needed. So, that's basically it.

[51:04 - 51:23] This is why we basically put the initial lecture as the orientation. It's partially basically for a couple reasons. One is we want to introduce you, basically, to all the tools and resources. It's not a traditional lecture. Basically, we basically have a coaching program and a community program, and then we want to introduce you to how to be successful, basically, as well.

[51:24 - 51:33] Michael asks, "How do we access the recorded session?" Basically, yeah, we're going to invite everyone, basically, very soon, but all recorded sessions will be in each unit.