🐙 Lunch & Learn: Building AI Products! | Tina Huang

0:00Hello friends. We are live. Yes. Yes, we are live. Could I get a sound check from anyone? Yeah, could we get a sound check for both of us right now? We are going to be a live in the apartment. I can't be the screen. Yeah, I know. We will always support you. Thank you. Thank you guys so much for joining. I'm with you on that. I do actually want to know how to do that.

1:00Can we watch it? Very interested. Tina, hello friends. Hey, Sophia. Hi, psych. Psych Shock. Elmar Maceto. Blogging roses. Hey, Christian. Hey, Artichwald. Eberhem is a little quiet. Okay. I only hear one person. It's audible. Mike sounds great. His audio is getting cut out a lot. The guy sounds robotic and clips away. Oh no, he breathed. Hello friends. Thank you all so much for joining. This is a really good turnout on live as well. So while Eberhem tries to fix his mic, I'm going to just kind of a little bit of, I'm going to share the slides and then just give a little bit of intro for what's happening.

2:01So are you okay? Eberhem? Oh yeah, actually, let's give an intro. So yeah, so that's Eberhem and I am me. So hello. Eberhem, do you want to give a brief intro about yourself? Yeah. Wait, oh yeah. You don't know what Eberhem looks like? Sound from you guys here, may I? Before I start, in 4-1 minute, just a little sleep. Okay, okay, sounds good. So hello everyone, it's my first time on Tina's chat. So I have been working at Porto. And I have recently, you know, I have a project to implement inside Logo octopus.

3:08Yeah, I'm very happy to be talking to everyone online. Oh my god, please. Oh yeah, crazy quiet. He boosted up a little bit. Still very quiet. Can't hear. No, video quality is so low. Sorry guys, we're trying this out for the first time. So this is already the highest. It goes apparently. Oh no. Well, I'll be sharing some slides so you can look at the slides and not look at us while we are pixelated things on screen. The video is stuck. His video, okay. Audio is breaching. Okay, Eberhem, I'ma just get started. Don't worry. It's okay, it's okay. We'll be okay. We have each other covered, so don't worry. But thank you guys all so much for joining. Also, let me know where you guys are joining from. I hope it's not in the middle of the night for everybody. And Afra, has this for beginners? Yes, it is for beginners.

4:09We did it so that beginners can understand, but we also included more things in that we can go into more details for anybody that has more advanced questions. But yes, it is for beginners. I'm going to share my screen and we will all hope that it works. Please work. Screen. Okay. Did it work? Okay. Let me, how do I make it not small? This is very hard. Oh, okay. Yes. Alright. Okay, so now you can ignore our pixelated faces so we can look at this. So I just actually want to give a context for what's happening here. So as part of the lonely activist team, I just wanted to thank you guys for making this possible for doing. So we wanted to do lunch and learns in which, like technically it's lunchtime and PSD time,

5:10although neither of us are there right now. But there are lunch and learns on different topics and data and AI. It's, these are completely free and it's just because we genuinely think that understanding about AI works, how data works, how they interact with each other and how it's changing in a data landscape. It's really important to know even just on a cursory level, I think it would help a lot. We hope it will help a lot in understanding what's happening and the things that you can potentially build out. So that's the purpose of the lunch and learns and we'll have a different guest with a different with like different speakers about different topics all related of course to like AI data tech career, that kind of stuff. So if you want to be notified for this, we do have like a, we don't want to call it, I'm just going to stick it in the chat and pin it. So you can either sign up for lonely activists if you're interested in the wait list, you can do that, but you can also do the email sign up.

6:12So either way is fine. Just in case you are interested, you can do that or you can do the email sign up either way, you will get notified. So email sign up to get notives and for future events plus summaries. So we're also going to send you guys summaries of things that are happening. Oh no, how do I edit it? Oh shit. Wait, let me do it again. Sorry. Ignore me. I'm notive and summary. You can sign up over here and if you want to do the joins love, I'll just leave the link over here, but not necessarily do so. Don't worry about it. Just leave me in here. How do I, okay, I apparently don't know how to do anything. That's not right. Oh shit.

7:13I did it again. Okay. Whatever. So lonely octopus. So this is the wait list if you guys want to do that. So that's going to be opening next week. Yes, it is being recorded. So it's going to stay here on YouTube. Do not worry. So I'm just going to pin the summaries if you want to just join that story. We're not going to send you anything that is outside these things. Cool. So I'm in London right now. Francis, we just did a meet up. Really? Alien vibes. Okay. Can you guys hear me right now? Is this the e-book? Yeah, it is the e-book that you know. That's what he looks like. Well, this be a weekly thing. Yeah, it's going to be like weekly to buy weekly. So we're planning it to be at the minimum by weekly, but we're scheduling people like next week. There's going to be people who are actually coming into do a workshop. They'll be doing a workshop on building an AR product. So it's going to be like different topics, but we at least buy weekly. It will be happening.

8:14All right. So I'm going to get started. So feel free to ask any questions, by the way. Yeah. So feel free to ask any questions throughout this in the chat. We also have a Q&A afterwards. So it's pretty chill. Okay. So here's the agenda. So first we're going to talk about when AI product is that we're going to talk about the AI type. So the main focus is going to be on AI product examples. So these are going to be, I can talk about like, oh, like this is like chain. We're like, oh, this is like the API. Right. How do you actually use these things in industry? So these are like actual projects in industry. So this is where we're going to highlight. Like we're going to, I think it will make a lot more sense about the things that we talk about earlier. And then we're going to end with a Q&A. So pretty chill. So let's first start off with like what is an AI product? Right? So honestly, it's like super simplest definition. It incorporates AI as a significant portion of its operation.

9:14So nothing fancy on that side. But I do want to like just specify if you want to be a little bit more specific. It's like, think about it as something a human could do with the ability to learn. So that's the first point. Something that a human could do with the ability to learn. A human is often necessary to kind of monitor the AI, but it should be able to do human functions. The second part is to be it can be customized for specific use cases. And the way that it's customized is through prompting. So instead of having a model and you having to fit stuff into that model, the AI part is able to be customized to fit your specific use cases. So let's see if there's any questions. What classifies a significant? See, like that's what I'm saying. Like it's not like very like clear. This is not actually let me give you an example of what I mean by this. So it's like something that creates the backbone that is that builds upon the learning aspect,

10:16builds upon the customized ability. It's like a big component of the product itself. So make me an example. So here's an AI versus non AI product. So AI content creation is like you get the AI to write the content, right? You monitor it. You let it like write it and you might give it feedback. But in the end, it's writing a content for you. This is a non AI product. It would be like automating a type of content that is being produced based on static parameters. Like if you're like, oh, I only want to like see content. I don't know, like you have a database, for example, and you're like, I want content about like what was there about like about SQL or like something like that. And it would just pull directly from there. It's like, and then it might have specific instructions about how it is that you want the content to be written as does that make sense? Like as opposed to an AI content creation, you would give it a prompt, like a generalized prompt, and it's going to be able to write it by itself as opposed to a non AI product where you have to specify like what it is that you want to be written as.

11:21And it's essentially like logic like if else statements in some ways. And even if you automate it, it doesn't necessarily make an AI content. So we'll talk about more like specifics about AI content, which I think would clear this up like AI products. So I would say like honestly don't worry too much about tailDR. It's basically something that learns and is adaptable to different needs. It just makes life a lot easier. And you're able to perform functionalities that you wouldn't be able to do so in the future. Let's see if there's any questions. Okay, so questions like fast question data sounds are suffering. I'm going to ask you guys to save that for Q&A. So anything like related to the thing itself, like please feel free to ask now. Okay, so Arctic ball, I'm going to show you when we show you the examples, I think that would be a lot clearer. So it's harder for me to explain abstractly on a high level.

12:22It's saying save you want to have a AI tutor, for example, which we'll go through, right? Because you're able to customize that to the level of the student. So it wouldn't just, it's like you're customizing it to their skill level the way that they learn things like that. Okay, so AI technologies, I'm going to go again go through this, which will see the implementation of just want you guys to have a high level overview. So AI technologies, honestly, like a lot of AI technologies, we're going to go through the core ones that are very fundamental. That's pretty much part of any AI product that you're building. The most, most, most simple thing is probably in the, like an LM API, plus some sort of prompt engineering, plus some type of UI. So this API, again, this is like not something that is like, oh, you can't have like other kinds of AI things like you can pull from or other types of large language models, but just for to say, it's simplicity and most general use cases, you have to have some sort of way for you to communicate with the large language model, what you do through prompt engineering.

13:23And then you have to have some sort of way in order for you to display that information, which is the user interface. Does that make sense? Oh, we're speaking of using a derivative chat, you be right. I suppose that training a model is not possible, we're not useful. So in this case, I just want to say like specifically we're pulling from GPT, like the large language model itself, while chativity is built as a chat bottle on top of that. So it's like there's a little bit of nuance to this. You can train a model, absolutely, you can train a model. Just in general, general sense, it's very rare in which you need to train a model by yourself. I guess just in most cases, there's no reason for that unless you have a privacy concerns with like very specific implementations for almost anything else, not necessary, and it would just increase your business cost. Okay, so more complex is that's the simplest, right? L-A-M-A-P-I-Plus-Promp Engineering-Plus-U-I.

14:25So, and then more complexity, usually you add in a database, and this is where you store predefined responses and parameters. This is really helpful for you firstly to cut costs, because you don't want to be constantly calling the large language model. Because every time you call, there's going to be a money that you have to pay in order to do this, which is not that rich, but if you scale it up to like industry level, it can really add up, and it's pretty significant. So, if you store things as predefined responses and parameters in a database, you only have to call the API when it's necessary to do so, when it's like you can't retrieve it from the database itself. And then there's also token limits, so when you're calling the large language model, there's a certain amount of tokens, where certain amount of things that you can put as part of your prompt. So, this limit is actually not that large, so this is a way for you to have these predefined things to bypass these token limits as well. That is the more complexity. Will the string be saved?

15:26Yes, it will be saved. Yes, it will be saved. Okay, so going on to even more complexity, so database for just caching? Not necessarily, it's a way for, it's more like you can predefined certain things. Like, database can be used in a variety of different ways, so you can predefined responses, you can pre- anything that you want to save essentially. So, anything you could do in order to not have to call the LOM API is generally the biggest use case for having something like this. Monitor, yeah, don't forget. Okay, cool. So, even more complexity, so this is where we have a large language model, you have API plus prompt engineering plus UI plus database, and then this is pain, which really is pretty much like part of a lot of projects. Now, this concept is part of this like technology, and the concept itself is super helpful

16:31in building AI products, which is chain. So, what chain is to be very simplified, it's a way of incorporating additional data among many other things, so that you're incorporating other data into the, what it is that you're building. So, it's not just calling from the large language model itself, you're incorporating maybe like your business specific data that you can use. The name comes from being, and so as well. So, if you have like something that you call the large language model and it gives you something else, then you can actually chain this. So, the thing that's outputted, you can again put it through to a large language model to get something else. So, this is a way that you can chain this without, without explicitly having to do so, so you can build out chains, hence the name chain. So, again, we're going to go into a lot more specificity. In fact, like for our projects, we start off, we actually show you the simplest use case for one of the sample projects, and then we'll show you what more complexity project looks like, and then what the even more complexity project looks like.

17:31So, it's going to make a lot more sense once we hit those parts. I did want to like briefly cover what that means. All right. So, this is the part where we actually get to the products itself. So, I just want to give like general reviews for each of these products first. So, this is actually, these are freelance projects. So, full credit goes to the people who create them, which Abraham will be talking about. So, freelance projects that were part of the only octopus. So, people who were interested, so they learned like data and AI skills, and then they applied them to actual freelance projects from companies. And this is where we like show, these are some of the AI projects. We also had some that were not, but we wanted to showcase like the AI ones, because that's what we're talking about right now. So, this one is a content creation project. It's for a healthcare clinic, and it's content creation for medical content. So, this is interesting because content creation for medical content means that you can't just say random things potentially. Right. Like, you have like hallucinations, they could actually impact people's lives.

18:35So, you want to make sure that your AI is just like writing articles about things that could be potentially harmful. Particularly important in this context. So, consideration of hallucinations in particular, and in this case, decline was interested in automatically posting to WordPress as a block. So, that's the type of content they were interested in doing. And the automation aspect was also important, because as a healthcare clinic, they have so many other things that are going on that they don't want to be, it's like having to like do it in a manual basis. So, let's see if you guys have any questions. What would be, okay, so, I'm going to go through this first, and I'm going to take it, you bring him, you want to take it away? Yeah, that is my voice. Hopefully. Yeah, you're fine. Hopefully. I think I can hear you fine. Oh, it seems to be fine. I'm in the live, it's a little weird.

19:35Maybe it doesn't take. It's a full credit. In John, it's Thomas, Italy. They'll take the LinkedIn, them some love. So, they have worked on this content creation, which would want to be the... Okay. First, yeah, it's not an interface. Really, you cannot be heard. You can't be heard right now. Can you, can you up your sound as poor from the dude? No, he's not understandable in my end. Shit. Yes, we can, I can hear you fine, Ibrahim. I think it's something. In this game.

20:38Sorry. Unfortunately, I'm in the end. Okay. Okay, I guess I'm going to do a moment. I got Ibrahim, you put him in the spot. Yeah. Oh, wait, try turning video off. Can you try turning your video off, youbrahim? I'm going to, I don't know, I guess I don't use her right off. Is this fine? It's really weird. It sounds like... Much better. Much better. Oh god. Much better. Okay, it sounds like it's... Talk faster. Talk faster?

21:41No, I mean, like quickly before it dies again. Okay, so, yeah, it's where you put the subjects on the target audience. You extract articles based on that from PubMed. And then you extracts into TPPT 3.5. And then the article gets. UI and then you have the option to into WordPress. Oh, wow, he has 1010 right now. Yeah, I'm always the issue. Okay, so okay. So, you know, turn my screen. I can do a live demo for them. And again, please check Junjin Thomas. Thomas is the LinkedIn. They've been awesome. Okay, that goes to them for watching on this amazing project. So, the screen so I can demo this for you. And hopefully, the effect of my screen is not going to be like

22:45the effect of my face. I think you can see it, too. You don't mute my voice, you know? Oh, okay. Yeah, I was going to say apparently turning my video off might help as well. So, I guess I will turn my video off as well. So, try. Oh, okay. I'm done. So, this is the application that has been developed inside. So, it's basically- Wait, Ibrahim, sorry. Can you make it a little bit bigger? It's hard to see. Okay. Okay, so- Yes. So, thank you. Okay. So, the application made inside all the octopus. So, the goal of this application is to increase traffic

23:47for wind-tenth through content creation. Like, maybe just write the topic. So, let's try aquaculture then back pain. Okay, on the second side. Check what- um, uh, yeah, and then you just press enter. And then you can- It says- uh, to try another topic of- puncture thing. Okay, so, we're going to- Again, it's about 12 articles. So, in the meantime, you can think about something. You can maybe- Good. In the spring time, it usually takes, uh, five-

24:51to generate that dependent on how big the article is. And the- uh, abstract, it has found. So, for- um, I mean, one way you can make this even cooler. These are the AI. But right now, we just have, uh, images and whenever an article gets generated, uh, it- uh, an image based on the topics right now, it- with back pain. And you can see the article is already- general- and- Yep, and now you can push it. You can push the draft to- when sends what's- I- and then you- click this- this is the part that drones mean. Go up into, uh, wordpress- and boom. You can see the article-

25:51generate and then you can edit it, and then you can publish. So, yeah. For- um, for this one. Uh, I'm gonna answer any questions. I- I hope I was heard- draw the presentation. Excuse me, let me know- it did not die. I mean, I hear you. I'm like very concerned that people didn't hear you. Is your screen blocked now? Very difficult to understand due to the dips. Might just mean me, but the stream quality is dipping. I believe his internet is bad. Okay, you know what? I'll do the- I'll do the demos. Oh wow. I- I- We- Look- Ah, but- but- but I- these don't-

26:51Yeah. Okay. What did you guys hear and not hear? Where you started, nothing? Let me send a link in the- and you can open it to see now, okay? So let me send a link in the demo. So- Yeah. You can click- Tina as well if you'd like. It's the link in the YouTube tab. I sent in the YouTube chat. The YouTube chat we're discording in. The YouTube tab? Yeah, I can't- I can't send it on- on Discord. So I have a gift, so- Oh, okay. So I'm just gonna share my screen, right? Okay. All right. We got this in- In Prop 2, and guys- In Prop 2. Okay. So screen. Okay. I will try my best. Okay. So in this case, as a writing assistant,

27:57are you able to see it? Yeah, okay, cool. As a writing assistant is helping you to draw a blog post based on a topic's in the audience provided. Draft will be based on recent publications and publics. So as we said, it's a health clinic. So we have to make sure that what is being written is actually based upon things that are published. So topics here is acupacier for Bunk. Oh, no, what happened? Okay. Yes. So at that after that was loading. So acupacier for back pain, as you saw earlier, that was- these are the articles that came out from it. So let's try that again. Acupacier for back pain. And who's target- so a dust workers generate article. Searching PubMed, and then we chill out for a bit. I just want to make sure people that this is like a rough live stream. Indeed. Indeed, it's a bit rough. Yeah. Do the-

28:57I can't. Okay. Maybe could you quickly view what he presented? Okay. So it's finding articles right now, searching PubMed, and then it generates a particular article. Acupacier for back pain, a fresh perspective for a desk workers. And as you can see, it would write through the articles, and then also provides the references. And then you can then post the draft onto WordPress. So it will automatically do that. And yay. It's here. Acupacier for back pain. With all the things that are written over here. So there might be things that you do want to change, right? So you want to make sure that you- maybe there's like certain things that you want to do differently. And then you might want to edit it yourself before posting. And as an extension of this project, this is like the MVP version. Something that I know the team was interested in working on is being able to showcase like several different articles. So our client is able to choose the one that they're most interested in.

29:58All right. Let's see if there's any questions. I'm going to- We'll see what answer all of them. Oh. The teams will answer all of them. Team. All right. Yeah, it seems to be the case that I will be answering them. Wait, it's okay. Let me just share my screen. The content creation. This is the demo that is here. So as you can see, we just- there's like a simple diagram that's being drawn. So you have the UI that's there that you saw. The PubMed abstracts are put into GPT 3.5. As you enter the subject and target audience, it generates the article, a UI displays it, and it posts the draft onto WordPress. So this is like a simplified show of doing of this. So as I was- remember how I was talking about earlier, like the simplest technology. So in this case is one of the simplest AI-based technologies, which is the large language model APIs, plus the prompt engineering, plus the UI. So this is an example.

30:59I do want to make an emphasis though about the fact that as you can see, the PubMed abstracts are not actually technically part of the core AI thing that you're doing, but because of the specific use case like this, it's very, very important. Actually, a lot of these technical details by figuring that out, like those end up being a lot harder to implement. Out. Then the actual AI part. It's self, which is linking to GPT 3.5. Also here, generating the article, and also being able to post the draft onto WordPress. So this is more like API-based stuff. There's API-based stuff over here. So I just want to make an emphasis. That this is what I mean when I said, like, oh, you can have the simplest kind of product, but since this is an actual product, that has a specific function for a business, it's important to incorporate those things into the product itself. It makes sense. See, people hear me. Sorry, but it's okay for you. Thank you for answering questions.

32:00Um, the Scrooping PubMed website are using API. I believe it is an API. If I'm correct. Yeah, yeah, it's the PubMed. Will the article pass AI content detector? What's AI content director? I don't know. Okay, so um. Oh no, I was like, I don't know what that is. Is that like a thing? Where is that just like the concept? No, no, I'm not aware of anything called that. Like, is it going to be like a proper article? Like, not the need for the information and so on? Yeah, I mean, the, the double-check the information. Yeah. Yeah, you should definitely even with PubMed references, you do not want to be just, okay, so that's also why it's as a draft instead of like directly posting like you want to be able to make sure that it works and that you're not posting

33:01anything like, oh, you're completely wrong. No problem, it happens. Thanks for the demo. Looks reasonable. Okay, so next product is AI. Product is the AI YouTube chatbot. So kind of give a little bit of overview about this project as well. So in this case, decline was a media agency and they wanted to help content creators reply to questions. Oops, I typed wrong. Apply to questions on the channel. So this can be done for any type of content. You know, that type of, so as content creators, one of the things is like we want to reply to questions on YouTube or like on any other content creation platforms. But it's oftentimes like there's a lot of, there's a lot of questions that are coming in and we just simply cannot manage it. It takes a lot of time to answer people's questions. And but we still want to do that. So this is a tool that can be used in order to answer those questions. And in a way that mimics the way that the content creator can answer the question. Also, I want to make an emphasis that it doesn't need to be like content creator. It doesn't need to be based upon like video transcripts

34:04on YouTube. It can be done in any type of content. Like maybe you have courses in which you want to be able to interact with a DChat bot that has access to as trained on the courses. And anything else that you, as long as it's like some type of content that is in text format, this would be able to do. So in this case, you want to make sure that's referencing specific videos. Similar to similar to the above to the like previous product. It's the fact that it's specific in this case that you want to be referencing specific videos to where it is that it's being pulled from. So you don't want it just to be saying like random things. And then also on this specific project, there's a lot of cost and time considerations. And you need to be able to answer the questions quickly. Right? Like you don't want to be like sifting through all the videos and doing all these things every single time. So, book, insertions are going to be there. Are you typing right now, Abraham?

35:05Uh-huh, that's how I think of it. Sorry, I'm going to get myself. Oh, okay, okay. Do you want to try again? No, no, no, you're going to cry. No. You don't want to try again. I don't want to try again. I'm going to send you the link again. And then so once again. Okay. Yeah, just to make sure to try. Yeah, so I send the link again. You can I'm thinking for the title. Okay, so this case, it was actually carrying on Kenji's videos. If you guys know who that is, so enter your query dangers of doing data science. And then being a data and data loads successfully. So these are the transcripts that were put in and the specific URLs of the YouTube videos that it was taking from. And then we let it run as it answers this question. And now you think about your existence. And the last. Yes.

36:06My audio does not go through. Okay. Yeah. And these are the answers that it was able to generate. And then you can also actually look at the videos themselves that was referencing that was being referenced to answer the questions. That's me. That's me. Okay. Cool. So going back to here, are there any questions? Okay. Okay, cool. All right. I'm assuming. Answering the questions. So building our product. So again, go show everybody here. Some love. The credit goes to them. So this is the team that worked on this specific.

37:11Yes, that worked on this specific project. So again, like a little, this is like a more complex project. So first we've got to have the YouTube playlist which is around 600 videos. We're so and then you have to extract the audio. YouTube videos. We have to turn that into audio formats through Pytube. And then we use a technology called DeepGram in order to transcribe this audio. Then we store this in a Postgres SQL database. So the video title, the links in the transcripts. And this goes through Lanching. This is in corporation of the data that's being stored in program Postgres. I hope that clarifies an example in which why you require a database to do this. And then on Lanching, which then provides the transcripts and those things would go into GPT3. And then it would formulate the answer. So you ask the question. It would formulate the answer. And then it displays on the UI. So as you can see, there's specifically, so all the things that are coming before that

38:13is again, very, very important to this project itself. And this is only, this is because of the certain requirements of the freelance project and of the client themselves. I need to watch the replay as I missed it. Don't worry, there is replay. I hope it's working. All I hear is silence. Okay, that's fine. Just talking to myself. Okay. All right. So next up is the AI product. It's called AI coding tutor. So kind of brief introduction to this. It's a data science. The client here was a data science platform to practice coding. So what they wanted was a tutor that gives hints for questions and teaches concepts to different students.

39:13I was saying earlier about the ability to conform or the ability to be able to adapt dynamically to different levels to like be able to customize it. This is an example of clearly what is considered an AI product. So it has to dynamically adjust to a level of different students and we need to make sure that it stays on topic. This is a really big one for this specific project. So if you just make a chat bot like a coding tutor and people are definitely going to go on a platform and ask it weird questions, like that is human nature for people to do that. And if asked these questions, we need to make sure that the bot doesn't just say something that is inappropriate, whereas it goes broke and answer questions in a bad manner. So we need to make sure it stays on topic. And as well as this project, there is a time and cost considerations as well. Because say like you're on the platform and somebody has a question where they want to hint for something and you don't want them to stay around

40:14for a very long period of time. That's poor user experience and also cost considerations. Since this platform has a significant number of active users, if we had to pull from the API constantly, that's just going to have a significant amount of cost. That's what we need to think about as well. Okay. Do I just click the post here? Do you bring him? Or are you going to say, ah, there you go. Okay, no, just kidding. Yes, just click the post. I'm like scared of clicking things, though. Okay. Ooh, this is from Tomioa. I was going to share this tab and says, so this is on LinkedIn. So she made a post, so give it some love, give the post some love. This is the link to the GitHub repo. And here is the demo.

41:15That's going to make that, try to make that a little bit bigger. Let's go through it again. So it's called strata AI. Very nice design. So first I want you to just want to emphasize that there's FAQ. So get walked through, show edge cases, explain solution, and seek further clarification. So we want these are the things that the client's specific that they wanted. So in this case, a question it could be how to sort a list. And then it's able to give this output. That's all coming over here. And then we can see the results. And you can interact with it as you're going through with it. Like after you do this, walk me through how to sort this. And then it's able to walk you through it. So you are and is able to answer these questions dynamically. Or if it's something that's predefined like FAQs or something like that, it's able to play, I can't talk.

42:15It's able to pull from a database instead of having to go and pull from the API directly. And as I was saying earlier, we need to make sure that it doesn't just go off topic. So we also tested the team tested how to make lunch. And the bot will say this question doesn't seem to be related to Python programming. Please provide a more specific Python related question, choose from the provider options, provide a relevant URL for context. So this is something that the team did have to really, keep bot factfully stayed on task. Okay. Aver any. Hello, please. Hello, Tina. Hello, Gangnam. Okay, cool. Aver, he's doing a great job. All right. So again, this is the structure of it.

43:17And this is what it goes to Matt for actually creating this diagram over here. So you have the user input, prompts, and input. It goes to the UI and goes to the GPT, a large language model back in over here. So also the database feeds into that. And then they also, there's a way of saving these conversations as well, which is really important if you want to review them. In case like there's something that you want to change, and then you also get the response back. So the response goes back to the UI that's being displayed. So this project I just want to say that this is the complexity. Like this is a good example of increased complexity because there are certain specific things that are needed on top of just the AI component of it. I can need a save conversation, needing to like put it on track. So these are all like infrastructural technical implementations that were necessary for these projects.

44:19So these things are things that you can't really do from what we call learning projects. Like you know how there's projects like, oh, like make a Twitter bot or something like that. I think those are like obviously great for learning things. But the reason why we always like highlight these freelance projects is because it really showcases how business implementation is very different than projects that you may be working on. My voice, I will use my words. By way of deception, I wish I could read that. I thought it really, really. I don't know. Okay, did I? Where maybe I'm just bad at talking? All right, so now we're going to open up for Q&A. So there were some questions that came made from Instagram and also through YouTube. So we're going to go through these right now.

45:20And feel free to ask any other questions that are that are interested in too. So I'm going to open up my Instagrams. I think I know how to do that. How can we use user feedback and real-life data to keep improving our AI? The Android app heard. Really good question. So user feedback is feeding it back into chat to be taught. So one of the things real-life data part is going to be, will we talk to you about using chain like incorporating additional data that's specific to a certain use case to improve your AI. And then with the user user feedback, for example, with the AI tutor example, see how they were saving the responses that are there. So the responses right now as a future implementation is the fact that you can take those responses, see if there's things that you want to be changed and feed it back

46:22as like a way of telling you like, oh, like these are the tweaks that you want to be made. And this can be done like directly through that, or you can be changing the prompt in order to incorporate things that you learn from user feedback. But 100% user feedback is a really big component of making sure that your the AI, like whatever this that you're building is in fact doing the things that you want it to do. That is one of the biggest challenges by using large language models is the fact that you don't want to start forgetting what the instructions were, which sometimes happens quite often, or it goes off-track because it is a generalized model. Right? So it can easily go off-track as well. So these are very important business considerations. What does one need in terms of resources and time? In terms of resource wise, you don't need too much. Like you need to buy access to one of the large language model APIs in the minimum.

47:25Like I think, with OpenAI, I'm not going to, I think it was like $10, $20 from us. So there's that. If you know how to code and if you have a database, you might want to store it on AWS or something like that. And in terms of time considerations, so these projects were completed in approximately two to four weeks, depending on the specific projects. And they were completed by people who are working full-time who are like, have additional things as well. So I just want to say, everybody that working these projects did have additional things that they were working on too. So if you actually just like sat there and just did it, all I want to go, probably like a week or so, but this case took around a two to four weeks to do. Let's see. How do you choose the right model and text that? So that is, it depends. I thought the best answer, but it depends a lot on your business implementation.

48:26Right? In terms of the text that generally, it's like, I think I have that slide. Let me just find a slide again. Yeah. So these are like kind of the core things that you almost always have for most AI products. And then there's other like aspects that go into it. Like maybe you want to make a website. Maybe then you start adding on different things. Yeah. So it's pretty much large language model API, prompt engineering, UI database, and chain. So these are like kind of the core things that people generally use. And then you got to add in different things, depending on the implementation of the project itself. Hi from France. How's it going? I got a question. So many questions.

49:27Do we really need AI into product? If you don't need AI into product, you don't need it per se, but I think generally by incorporating some AI things into a product, which is it helps a lot of optimizing certain operations. Like some of my friends that still work in big tech companies and we chat about this quite a lot. Like they don't really come up with particular AI products. But the tooling and the infrastructure and the things that are currently there, they're far better now in doing what it is that they need to do. And that's by incorporating AI into it. What should be my next step for making chatbot if I know Python only? So I mean, there's the one chatbot that we showcased here, but like a even simpler example, instead of incorporating like additional data into it, you can just build off like chativity,

50:27like that is the chatbot. So you can use that. And then you can like tweak it by giving a specific questions. Like for example, you can get sorry, specific instructions like, you know, you can just in general tell it, oh, like can you speak in a certain way, or can you act like a certain character? So I would say that's an even easier implementation without having any external data that involves. Insights and building products are my idea to end the day to a product that people actually use. So that's yeah, that's another thing. Like if you start with the problem itself, like these projects were designed to solve a specific problem from a freelance client. And I don't think you can really come up with this kind of thing just by yourself, right? It's like it's hard to come up with this if you don't actually have a problem to solve.

51:28Qradia events also online. Yes, they're also online. How do you stay up to date without the AI tools coming out? So I think that's a really great question. There are so many AI tools that are coming out all the time, do I keep up with all of them? No, I don't. I think the what's important is going back to like the core component is like understanding how AI works and understanding from an engineering side as well, like how especially how databases work and also how APIs work. It's just it's like all the other tools become a lot easier. And you find out about the tools as you're building something, you're like, oh, like I need, I don't know, like the PubMed API array. So then go and look at the PubMed API. To the power, yes, I can connect it through this. So I think learning the fundamentals about data, it really opens up like such that you don't need to know every single thing that's coming out because you're able to use

52:30everything. That's already that's available. It looks that makes sense. So yeah, I would say the most fundamental skills, I really recommend people knowing. So first one is how APIs work, especially in an access question, but I've had this discussion with many people about how it is that the data field is occurring right now. And I would say a really big component is that data scientists, data people in general are much more expected to lean on engineering principles like how to use APIs and things like that. A simple analytic space stuff is just not really cutting it anymore, especially because so many of these large language models are able to analyze the data, right? So it's like being able to use this is extremely important. And if going hand in hand with that would be prompt engineering, extremely important because that's how you interact with the large language models, like figuring out how to prompt properly can make a world of difference in the product that you're building.

53:35And another one would be the third, like the next thing would just be databases. I think it's extremely important to understand how databases work. Because they're just very, very common, like the incorporation of external data into a large language model into doing these things, it's very, very common to do that. Sorry, I'm going to go. Thank you again for doing this to both of you. Thank you for joining. Okay. And let's see. Other questions? Question. I don't even know what is an AI product. Well, hopefully you know what an AI product is now. And I'm working on my first project. I'm lost in the how do I know what I need sauce. Okay, there you go. I just told you. Perfect. Two birds one stone. I just told you those are the things that you should know. And to create a customized chatbot is lane chain and open API.

54:40The way to go is there a better approach. I think it's the easiest implementation right now, like especially because both these tools are, it's like very well documented. Probably documented by AI. Is really well documented. Are there any other approaches? Absolutely. And I do think it's probably the most well documented. Any future AI big project you're working on? So as we're opening up, Lonely Octopus next week for the next iteration and we're opening up applications, we do have some pretty cool projects from other businesses that are coming out. Since those are like AI products that we're adding on to. So it would be incorporating AI that's based upon optimizing AI products. Things like that. Those things would be coming out. Okay. Is there any like future questions?

55:42YouTube guide. You could give your thoughts on AI alignment. Is it not being done in the industry to ensure AI is safe and accurate to use where it's not being lost in the hype? Truthfully, I think. No. Yeah. Oh. As you can see, like even from these projects that are being built, there's like you kind of have to think about what potentially could go wrong. For example, making sure that content is being produced but sourcing it but making sure that there's sources. I can use like didn't do that. Then, you know, potentially there's things that are going to be encouraged.

56:43And I think I would hazard, I guess, a lot of products that are being built now don't just really consider these things because it's not necessary to get it to work. I need to come up with my ideas. Do I need to master linear algebra first where is there a shallow end to the pool? Well, I would say the most shallow end, I would highly recommend that you do linear algebra and I highly recommend that you do stats at the minimum. But technically, technically, in order to just use the API and then build it up more of like an engineering kind of thing so you don't technically need math but it would be very helpful to do so. So you'd be able to understand how large

57:44language models work under the hood. An AIB creative without using mathematics? No, it is very much needs math. And I'd be creative using AI. I don't see why not. And I haven't tried it but I don't see why not. It could not be done so. What code in language is best for AI as a beginner? I would say it's like what I say in general purpose why I just Python is really easy language to learn and it's just very multifunctional. I can we end with one percent? Tina is if you're ending now. I want to know what they want to build with the AI. What things? After saying this presentation, what do they have in mind? What kind of product or application they want to do?

58:45Thank you for delaying my message. Thank you. Sounded like a robot. Okay, well relaying Iberheans final question. So yeah, after seeing this presentation, what are some things that you guys are interested in building? Also write down if there's anything specific that you're interested in learning about in the next couple of luncheon learns. So as I mentioned, the other luncheon learns are going to be the one after this is going to be hosted so it will be people who are all actually as meta AI people? Yeah, I think they're actually all as meta AI people and they'll be doing a workshop on how to build a product like an AI product.

59:52Oh wait, yeah, I'm just going to pin this. I'm going to pin that. Okay, did you go to London? How is the meat? It was really good. I've put some photos in the denominator. I so screwed up though. Like I literally got the date wrong and then I got the I got the date wrong. I felt like a hard part, but I didn't realize that high part is gary large and there isn't like an actual main entrance. I feel so bad if anybody did come, which I think they're probably where people became and then they got lost because they didn't know which entrance I was talking about. So I'm so sorry if that was you. Even if people who showed up there like super confused about it. How long has this I missed the whole date? Yeah, it's either you missed the whole thing. I'm sorry. Okay, beside an IT skills for 2024 to future improve our career. Is that what you're building or is that a question?

1:00:53Oh, okay. I think it's a question. Never mind. Okay, so again, like I think being able to understand I just think in general there is a much bigger emphasis on understanding engineering principles. Even if you're a student with a data analyst or a data scientist, there is that necessity now of also understanding data engineering concepts, understanding how to build things from engineering principles. So that's where it's headed. House plan generator. House plan generator. If I can put in some kind of language, respecting constraint. What's all like designing houses, like architecture, that's so cool. I read it as house plan. I'm like designing house plans. Respecting constraints in a core to existing plan. That's pretty cool. Wait, have you tried, have you tried mid-Jerny for that? I was talking to an architect who was using mid-Jerny

1:01:55to take the sketches that he had. And then creating plans out of it, especially was like waiting it in a way. Like that was, he was able to come up with pretty cool. Like, pretty cool. What is that called? Drawings? Draw. Yeah. I'm going to build an LQ engine who runs such control stuff. Control what? Wait, and what is, what is it all? I'm assuming that we're talking about how to say is it not the language processing? Oh, okay. Battery. Undecure a subset. Relsead for systems closures馬上會

1:02:57Sorry, I just needed to look that up. Are you planning to go backwards and check in to call out corporations like Google Twitter? No, at least not currently. I'll look at that direction. Checking data for a while to be put into your system would be the most important task. Alright, we're going to end the lunch and learn now, so I hope that was helpful. So it's good on the timing. It was an hour and six minutes. So put in right now, please do put in Indy Chat, anything you're particularly interested in learning. And I'll put a poll up as well on Instagram. So yeah, if you're not following me Instagram, please do that.

1:04:01So that's where I give like updates, ask people for questions and stuff. And I give updates a lot more than I do on YouTube. So I'm just going to look at that here. So please add me on Instagram. Thank you so much for joining me. And yeah, that will write correct engineering documents for me. Alright, when is your next study with Tina? No, no, I would love to know. I think it's like mostly the timing thing, but I will. I'll schedule it and let you guys know. I'm sorry. You're like, I'm a pissant. Thanks for the stream from Berlin. Thank you. Alright. We shall be, I'll be ending this stream now. Thank you all so much for joining. And I'll see you guys in the next lunch and learn. Or video for livestream.

1:05:08You