[0:02] AI agents
[0:07] AI. Since the majority of videos on a
[0:10] YouTube about AI either too technical or
[0:13] too basic, I decided to record this one
[0:16] for all of our subscribers who either
[0:19] have just a little bit or no technical
[0:21] background at all. In this video, I will
[0:24] explain what AI, AI workflows and AI
[0:28] agents are. But don't worry, I will take
[0:31] it step by step and explain it in a
[0:34] simple words and we'll utilize our
[0:35] educational strategy we use in our boot
[0:38] camp starting with the things that you
[0:40] are already familiar with just like chat
[0:43] GBT rock or cloud apps. Then we're going
[0:46] to move forward to AI workflows and
[0:48] finally to the most popular hype words
[0:51] these days such as AI agents. And I
[0:54] promise those intimidating words such as
[0:57] rag, react or agents are going to become
[1:00] much simpler once we break them down
[1:03] using real life examples. But before we
[1:06] get started, let me quickly remind you
[1:08] who we are and then we will kick it off.
[1:10] My name is Sergio Kchenko. I'm a
[1:11] software QA engineer, lead manager, and
[1:13] a senior engine manager of ASD that in
[1:15] the past, but these days I'm helping
[1:18] people like you to become a QA from
[1:20] scratch or to improve your existing
[1:22] skills. Now, let's get
[1:25] [Music]
[1:29] started. Level one, large language
[1:32] models or LLMs. Every single app you
[1:36] have played with in terms of AI such as
[1:39] Chat GPT, Google's Gemony, Claude or
[1:42] Tesla's Oron Musk's Grog, those are just
[1:46] applications built on a top of LLMs. And
[1:50] LLMs are simply brains for every single
[1:53] AI app. Let me explain that in a simple
[1:56] words. Imagine using Google. You type
[1:58] something, Google looks it up, you get
[2:01] an answer. So that's pretty much how
[2:03] nowadays internet works, right? In a
[2:05] simple words. Now let's think about
[2:07] LLMs. You give a prompt or same input.
[2:10] You pretty much type something. Then it
[2:13] uses what it learned and it will get you
[2:15] the response just like in a Google but
[2:18] instead of the Google's server, we have
[2:20] LLMs or large language model which uses
[2:24] a lot of data behind it that was it was
[2:26] trained on. So far pretty
[2:28] straightforward, right? input LLM and
[2:31] your response. Easy peasy. But here's
[2:33] something to keep in mind. Every single
[2:36] LLM or large language model was trained,
[2:39] even though it was trained on a lot of
[2:41] data, it doesn't know your preferences.
[2:44] It doesn't know much information about
[2:46] you. It cannot access your files, your
[2:48] order history, or realtime info unless
[2:51] you specifically connected to it. And
[2:54] most importantly, LLMs are passive,
[2:56] which means they're simply sitting there
[2:58] and waiting for you to input some data
[3:01] to ask them questions. They don't go and
[3:04] do things on their own. And that's why
[3:07] we need
[3:09] workflows. Now, let's imagine familiar
[3:12] situation. You are hungry and you're
[3:15] telling your AI or chat GPD, next time
[3:19] I'm going to be talking about food.
[3:21] offer me something based on my previous
[3:23] order history from Uber Eats and it will
[3:26] definitely fail because it has no idea
[3:29] what Uber Eats are in terms of your
[3:31] orders cuz you did not give access to
[3:34] Uber Eats. But if you will build a
[3:37] workflow, then you could give access to
[3:41] chat GPT or the other LLM app. So it
[3:45] could access your orders previously,
[3:47] take a look at them, and then give you a
[3:49] recommendation based on your previous
[3:51] orders. And you could even specify such
[3:53] as tabs as check past orders, fix
[3:56] something with a five plus star rating,
[3:59] send a suggestion. And that works until
[4:02] you say this. Can you recommend
[4:04] something new and then you will fail?
[4:07] Why? Because the path you programmed
[4:10] doesn't include look for new options.
[4:12] Workflows are just like recipes. They
[4:15] only work the way they were written. And
[4:18] in technical terms, it's called control
[4:21] logic. Now, let's say you expand it a
[4:24] little bit. You add Yelp reviews, Google
[4:27] Maps, and make it read the
[4:29] recommendation out loud.
[4:32] Great. Based on your past preferences, I
[4:34] found a five-star vegan sushi place
[4:36] nearby. Want directions? Still just a
[4:39] workflow. Still not really thinking, but
[4:42] doing step by step what you told it to
[4:45] do. Quick myth or technical term
[4:48] breaker. You've heard about term rag,
[4:50] which is retrieval, augmented
[4:52] generation. Don't worry, it's just a
[4:54] fancy way of saying look something up
[4:57] before answering. Just the way we do
[5:00] with the workflows. That's it. It's just
[5:02] one of the tools inside of the workflow.
[5:05] And here is actually the main problem of
[5:07] workflows. If the output isn't right, if
[5:11] the recommendation doesn't sound tasty,
[5:14] you will have to go back, tweak the
[5:17] setup, and run it again. You are still
[5:19] doing thinking. You're still the one in
[5:22] control. And that's where AI agents come
[5:25] [Music]
[5:27] in. Let's stay with the same example.
[5:30] Right now you're the one making
[5:32] decision, what tools to connect, what
[5:34] steps to follow, when to make changes.
[5:37] But what if we replace you a human
[5:40] decision maker with actual AI itself?
[5:43] Then that's the time when workflow,
[5:46] fancy workflow becomes even fancier AI
[5:49] agent. And an agent can do three things
[5:53] that workflow cannot. Number one, reason
[5:57] or reasoning. It doesn't simply check
[6:00] the past orders. It thinks, "Oh, it's
[6:02] actually cold outside today. Maybe he
[6:04] wants something warm." Or, "Oh, he did
[6:08] order sushi 20 times within the last 10
[6:10] days. Probably we should offer him
[6:12] something else." It reasons through
[6:15] context just like you would. Number two,
[6:19] act. Now, AI says, "Let me check if
[6:22] there are any restaurants nearby." Then
[6:24] it will also check if they're actually
[6:27] open now. And finally, it will probably
[6:29] read some reviews before making
[6:31] decision. And it does that all on its
[6:34] own without you giving exact stepbystep
[6:38] instruction just like you did with the
[6:40] workflow. It figured it out on its own.
[6:43] Let's say the agent suggests a
[6:45] restaurant, but then realizes the
[6:48] reviews are meh. Maybe the curry was too
[6:51] spicy. It doesn't send you
[6:53] recommendation and simply call it a day.
[6:56] What it does, it goes back, finds a
[6:59] better option and only then sends you
[7:02] final and improved result. That's called
[7:05] iteration. And it can actually do
[7:06] something cooler. It can critique its
[7:09] own answer with another AI model before
[7:12] even sending it to you. That's what
[7:14] makes it different from regular
[7:16] automation or workflow. It can actually
[7:19] think. Second nerdy word of the day,
[7:22] react. You might have heard this word
[7:24] but probably got scared because it's
[7:26] something related to AI that you haven't
[7:28] heard of. But let's take it apart. It's
[7:30] very simple. Re stands for reason and
[7:34] act stands for acting just like the
[7:37] second difference with the workflow. And
[7:40] that's why React is one of the world's
[7:42] most popular frameworks for building AI
[7:46] agents. And that's what pretty much it
[7:48] does, right? aside of iterating. By the
[7:51] way, I've been mentoring people for 8
[7:53] years. And if you're interested to learn
[7:54] how to build AI workflow or AI agent or
[7:57] how to become a software QA engineer
[7:59] completely from scratch, you can follow
[8:01] the links right below this video. Let's
[8:05] continue. Let's take a look at the real
[8:07] life examples. And we have a lot of
[8:09] different ones, but I'm going to pick
[8:11] one of the most useful ones for our
[8:13] community. Detecting wildfires. I'm
[8:15] going to click right here. And by the
[8:16] way, we're using landing AI website. And
[8:20] I'm going to click on one of the
[8:22] pictures to see how exactly it will
[8:26] detect the wildfire because we cannot
[8:28] have people all over the world
[8:31] constantly monitoring all of the forest
[8:33] cameras simultaneously. But we can have
[8:35] AI and you can clearly see it right here
[8:38] that there is a fire. As the human you
[8:40] can see that but AI can also see that
[8:43] and you can even specify the threshold
[8:45] of how confident it should be. If we
[8:47] minimize the threshold it will specify
[8:50] more you will find more fires or more
[8:53] things that look like fire. If we
[8:56] increase it then it will be more
[8:58] challenging. Let's say that way it will
[9:00] be more sure before it say hey there is
[9:04] a wildfire. So if you if you go all the
[9:06] way to 100% it will say that there is
[9:08] none. But if you at least get it to 85
[9:11] or 90, you will still show you, hey,
[9:13] that there is a fire and the government
[9:16] or any companies that run this AI can
[9:19] simply specify this threshold. Now,
[9:22] quick recap. Level one, LLMs. It's just
[9:25] like human brains. You ask it a
[9:28] question, it gets you an answer, just
[9:30] like the thing that we have right here.
[9:32] Level two, AI workflows. You design a
[9:35] process and AI follows it step by step.
[9:38] And level three, AI agents. You give it
[9:42] a goal and AI will figure out how to get
[9:45] there on its own. If this video helped
[9:48] you to make sense out of AI, AI
[9:51] workflows and AI agents, and now you're
[9:53] thinking possibly to build one. Leave a
[9:56] comment right below this video and let
[9:57] me know what would you like me to show
[9:59] you how to build on the next video.
[10:01] either send a message create AI
[10:03] workflows or create AI agent and I will
[10:07] do it on the next video. Now, thank you
[10:09] for watching. Get some workout, get some
[10:11] rest, drink a lot of water and I'll see
[10:12] you next time.