AI Agents, Clearly Explained — Transcript

Full transcript

1. 0:03AI. AI. AI. AI. AI.
2. 0:07AI. You know, more agentic. Agentic
3. 0:10capabilities. An AI agent. Agents.
4. 0:12Agentic workflows. Agents. Agents.
5. 0:15Agent. Agent. Agent. Agent. Agentic.
6. 0:19All right. Most explanations of AI
7. 0:20agents is either too technical or too
8. 0:23basic. This video is meant for people
9. 0:26like myself. You have zero technical
10. 0:28background, but you use AI tools
11. 0:30regularly and you want to learn just
12. 0:33enough about AI agents to see how it
13. 0:36affects you. In this video, we'll follow
14. 0:38a simple one, two, three learning path
15. 0:41by building on concepts you already
16. 0:43understand like chatbt and then moving
17. 0:46on to AI workflows and then finally AI
18. 0:49agents. All the while using examples you
19. 0:52will actually encounter in real life.
20. 0:55And believe me when I tell you those
21. 0:56intimidating terms you see everywhere
22. 0:58like rag, rag, or react, they're a lot
23. 1:02simpler than you think. Let's get
24. 1:04started. Kicking things off at level
25. 1:05one, large language models. Popular AI
26. 1:08chatbots like CHBT, Google Gemini, and
27. 1:10Claude are applications built on top of
28. 1:14large language models, LLMs, and they're
29. 1:17fantastic at generating and editing
30. 1:19text. Here's a simple visualization.
31. 1:21You, the human, provides an input and
32. 1:24the LLM produces an output based on its
33. 1:27training data. For example, if I were to
34. 1:29ask Chachi BT to draft an email
35. 1:31requesting a coffee chat, my prompt is
36. 1:33the input and the resulting email that's
37. 1:36way more polite than I would ever be in
38. 1:37real life is the output. So far so good,
39. 1:40right? Simple stuff. But what if I asked
40. 1:43Chachi BT when my next coffee chat is?
41. 1:47Even without seeing the response, both
42. 1:49you and I know Chachi PT is gonna fail
43. 1:52because it doesn't know that
44. 1:53information. It doesn't have access to
45. 1:56my calendar. This highlights two key
46. 1:58traits of large language models. First,
47. 2:00despite being trained on vast amounts of
48. 2:02data, they have limited knowledge of
49. 2:04proprietary information like our
50. 2:07personal information or internal company
51. 2:09data. Second, LLMs are passive. They
52. 2:12wait for our prompt and then respond.
53. 2:14Right? Keep these two traits in mind
54. 2:17moving forward. Moving to level two, AI
55. 2:19workflows. Let's build on our example.
56. 2:21What if I, a human, told the LM, "Every
57. 2:25time I ask about a personal event,
58. 2:26perform a search query and fetch data
59. 2:29from my Google calendar before providing
60. 2:31a response." With this logic
61. 2:33implemented, the next time I ask, "When
62. 2:35is my coffee chat with Elon Husky?" I'll
63. 2:38get the correct answer because the LLM
64. 2:40will now first go into my Google
65. 2:42calendar to find that information. But
66. 2:45here's where it gets tricky. What if my
67. 2:48next follow-up question is, "What will
68. 2:50the weather be like that day?" The LM
69. 2:53will now fail at answering the query
70. 2:55because the path we told the LM to
71. 2:57follow is to always search my Google
72. 3:00calendar, which does not have
73. 3:02information about the weather. This is a
74. 3:04fundamental trait of AI workflows. They
75. 3:07can only follow predefined paths set by
76. 3:10humans. And if you want to get
77. 3:12technical, this path is also called the
78. 3:15control logic. Pushing my example
79. 3:17further, what if I added more steps into
80. 3:20the workflow by allowing the LM to
81. 3:22access the weather via an API and then
82. 3:24just for fun use a text to audio model
83. 3:26to speak the answer. The weather
84. 3:28forecast for seeing Elon Husky is sunny
85. 3:31with a chance of being a good boy.
86. 3:33Here's the thing. No matter how many
87. 3:35steps we add, this is still just an AI
88. 3:39workflow. Even if there were hundreds or
89. 3:41thousands of steps, if a human is the
90. 3:44decision maker, there is no AI agent
91. 3:47involvement. Pro tip: retrieval
92. 3:49augmented generation or rag is a fancy
93. 3:52term that's thrown around a lot. In
94. 3:54simple terms, rag is a process that
95. 3:56helps AI models look things up before
96. 3:58they answer, like accessing my calendar
97. 4:00or the weather service. Essentially, Rag
98. 4:03is just a type of AI workflow. By the
99. 4:06way, I have a free AI toolkit that cuts
100. 4:07through the noise and helps you master
101. 4:09essential AI tools and workflows. I'll
102. 4:10leave a link to that down below. Here's
103. 4:12a real world example. Following Helena
104. 4:14Louu's amazing tutorial, I created a
105. 4:17simple AI workflow using make.com. Here
106. 4:19you can see that first I'm using Google
107. 4:21Sheets to do something. Specifically,
108. 4:23I'm compiling links to news articles in
109. 4:25a Google sheet. And this is that Google
110. 4:28sheet. Second, I'm using Perplexity to
111. 4:31summarize those news articles. Then
112. 4:34using Claude and using a prompt that I
113. 4:36wrote, I'm asking Claude to draft a
114. 4:38LinkedIn and Instagram post. Finally, I
115. 4:42can schedule this to run automatically
116. 4:44every day at 8 a.m. As you can see, this
117. 4:46is an AI workflow because it follows a
118. 4:49predefined path set by me. Step one, you
119. 4:52do this. Step two, you do this. Step
120. 4:55three, you do this. And finally,
121. 4:57remember to run daily at 8 am. One last
122. 4:59thing, if I test this workflow and I
123. 5:02don't like the final output of the
124. 5:05LinkedIn post, for example, as you can
125. 5:08see right here, uh, it's not funny
126. 5:10enough and I'm naturally hilarious,
127. 5:11right? I'd have to manually go back and
128. 5:16rewrite the prompt for Claude. Okay? And
129. 5:20this trial and error iteration is
130. 5:23currently being done by me, a human. So
131. 5:25keep that in mind moving forward. All
132. 5:27right, level three, AI agents.
133. 5:29Continuing the make.com example, let's
134. 5:31break down what I've been doing so far
135. 5:33as the human decision maker. With the
136. 5:36goal of creating social media posts
137. 5:37based off of news articles, I need to do
138. 5:39two things. First, reason or think about
139. 5:43the best approach. I need to first
140. 5:44compile the news articles, then
141. 5:46summarize them, then write the final
142. 5:48posts. Second, take action using tools.
143. 5:51I need to find and link to those news
144. 5:53articles in Google Sheets. Use
145. 5:55Perplexity for real-time summarization
146. 5:58and then claw for copyrightiting. So,
147. 6:00and this is the most important sentence
148. 6:01in this entire video. The one massive
149. 6:04change that has to happen in order for
150. 6:06this AI workflow to become an AI agent
151. 6:09is for me, the human decision maker, to
152. 6:13be replaced by an LLM. In other words,
153. 6:16the AI agent must reason. What's the
154. 6:19most efficient way to compile these news
155. 6:20articles? Should I copy and paste each
156. 6:22article into a word document? No, it's
157. 6:24probably easier to compile links to
158. 6:26those articles and then use another tool
159. 6:28to fetch the data. Yes, that makes more
160. 6:30sense. The AI agent must act, aka do
161. 6:34things via tools. Should I use Microsoft
162. 6:37Word to compile links? No. Inserting
163. 6:39links directly into rows is way more
164. 6:41efficient. What about Excel? M. So the
165. 6:44user has already connected their Google
166. 6:45account with make.com. So Google Sheets
167. 6:47is a better option. Pro tip. Because of
168. 6:49this, the most common configuration for
169. 6:51AI agents is the react framework. All AI
170. 6:55agents must reason and act. So
171. 6:59react. Sounds simple once we break it
172. 7:01down, right? A third key trait of AI
173. 7:03agents is their ability to iterate.
174. 7:06Remember when I had to manually rewrite
175. 7:08the prompt to make the LinkedIn post
176. 7:10funnier? I, the human, probably need to
177. 7:13repeat this iterative process a few
178. 7:15times to get something I'm happy with,
179. 7:17right? An AI agent will be able to do
180. 7:19the same thing autonomously. In our
181. 7:22example, the AI agent would autonomously
182. 7:25add in another LM to critique its own
183. 7:28output. Okay, I've drafted V1 of a
184. 7:30LinkedIn post. How do I make sure it's
185. 7:32good? Oh, I know. I'll add another step
186. 7:34where an LM will critique the post based
187. 7:36on LinkedIn best practices. And let's
188. 7:38repeat this until the best practices
189. 7:40criteria are all met. And after a few
190. 7:42cycles of that, we have the final
191. 7:45output. That was a hypothetical example.
192. 7:47So let's move on to a real world AI
193. 7:50agent example. Andrew is a preeeminent
194. 7:53figure in AI and he created this demo
195. 7:55website that illustrates how an AI agent
196. 7:58works. I'll link the full video down
197. 8:00below, but when I search for a keyword
198. 8:02like skier, enter the AI vision agent in
199. 8:07the background is first reasoning what a
200. 8:10skier looks like. A person on skis going
201. 8:12really fast in snow, for example, right?
202. 8:14I'm not sure. And then it's acting by
203. 8:18looking at clips in video footage,
204. 8:22trying to identify what it thinks a
205. 8:24skier is, indexing that clip, and then
206. 8:29returning that clip to us. Although this
207. 8:32might not feel impressive, remember that
208. 8:34an AI agent did all that instead of a
209. 8:36human reviewing the footage beforehand,
210. 8:39manually identifying the skier, and
211. 8:42adding tags like skier, mountain, ski,
212. 8:45snow. The programming is obviously a lot
213. 8:47more technical and complicated than what
214. 8:49we see in the front end, but that's the
215. 8:51point of this demo, right? The average
216. 8:53user like myself wants a simple app that
217. 8:56just works without me having to
218. 8:58understand what's going on in the back
219. 9:00end. Speaking of examples, I'm also
220. 9:02building my very own basic AI agent
221. 9:05using Nan. So, let me know in the
222. 9:07comments what type of AI agent you'd
223. 9:08like me to make a tutorial on next. To
224. 9:11wrap up, here's a simplified
225. 9:12visualization of the three levels we
226. 9:14covered today. Level one, we provide an
227. 9:17input and the LM responds with an
228. 9:19output. Easy. Level two, for AI
229. 9:22workflows, we provide an input and tell
230. 9:24the LM to follow a predefined path that
231. 9:27may involve in retrieving information
232. 9:29from external tools. The key trait here
233. 9:31is that the human programs a path for LM
234. 9:34to follow. Level three, the AI agent
235. 9:37receives a goal and the LM performs
236. 9:39reasoning to determine how best to
237. 9:41achieve the goal, takes action using
238. 9:44tools to produce an interim result,
239. 9:46observes that interim result, and
240. 9:48decides whether iterations are required,
241. 9:51and produces a final output that
242. 9:53achieves the initial goal. The key trait
243. 9:56here is that the LLM is a decision maker
244. 9:58in the workflow. If you found this
245. 10:00helpful, you might want to learn how to
246. 10:02build a prompts database in Notion. See
247. 10:04you on the next video. In the
248. 10:05meantime, have a great one.