0:07
AI. You know, more agentic. Agentic
0:10
capabilities. An AI agent. Agents.
0:12
Agentic workflows. Agents. Agents.
0:15
Agent. Agent. Agent. Agent. Agentic.
0:19
All right. Most explanations of AI
0:20
agents is either too technical or too
0:23
basic. This video is meant for people
0:26
like myself. You have zero technical
0:28
background, but you use AI tools
0:30
regularly and you want to learn just
0:33
enough about AI agents to see how it
0:36
affects you. In this video, we'll follow
0:38
a simple one, two, three learning path
0:41
by building on concepts you already
0:43
understand like chatbt and then moving
0:46
on to AI workflows and then finally AI
0:49
agents. All the while using examples you
0:52
will actually encounter in real life.
0:55
And believe me when I tell you those
0:56
intimidating terms you see everywhere
0:58
like rag, rag, or react, they're a lot
1:02
simpler than you think. Let's get
1:04
started. Kicking things off at level
1:05
one, large language models. Popular AI
1:08
chatbots like CHBT, Google Gemini, and
1:10
Claude are applications built on top of
1:14
large language models, LLMs, and they're
1:17
fantastic at generating and editing
1:19
text. Here's a simple visualization.
1:21
You, the human, provides an input and
1:24
the LLM produces an output based on its
1:27
training data. For example, if I were to
1:29
ask Chachi BT to draft an email
1:31
requesting a coffee chat, my prompt is
1:33
the input and the resulting email that's
1:36
way more polite than I would ever be in
1:37
real life is the output. So far so good,
1:40
right? Simple stuff. But what if I asked
1:43
Chachi BT when my next coffee chat is?
1:47
Even without seeing the response, both
1:49
you and I know Chachi PT is gonna fail
1:52
because it doesn't know that
1:53
information. It doesn't have access to
1:56
my calendar. This highlights two key
1:58
traits of large language models. First,
2:00
despite being trained on vast amounts of
2:02
data, they have limited knowledge of
2:04
proprietary information like our
2:07
personal information or internal company
2:09
data. Second, LLMs are passive. They
2:12
wait for our prompt and then respond.
2:14
Right? Keep these two traits in mind
2:17
moving forward. Moving to level two, AI
2:19
workflows. Let's build on our example.
2:21
What if I, a human, told the LM, "Every
2:25
time I ask about a personal event,
2:26
perform a search query and fetch data
2:29
from my Google calendar before providing
2:31
a response." With this logic
2:33
implemented, the next time I ask, "When
2:35
is my coffee chat with Elon Husky?" I'll
2:38
get the correct answer because the LLM
2:40
will now first go into my Google
2:42
calendar to find that information. But
2:45
here's where it gets tricky. What if my
2:48
next follow-up question is, "What will
2:50
the weather be like that day?" The LM
2:53
will now fail at answering the query
2:55
because the path we told the LM to
2:57
follow is to always search my Google
3:00
calendar, which does not have
3:02
information about the weather. This is a
3:04
fundamental trait of AI workflows. They
3:07
can only follow predefined paths set by
3:10
humans. And if you want to get
3:12
technical, this path is also called the
3:15
control logic. Pushing my example
3:17
further, what if I added more steps into
3:20
the workflow by allowing the LM to
3:22
access the weather via an API and then
3:24
just for fun use a text to audio model
3:26
to speak the answer. The weather
3:28
forecast for seeing Elon Husky is sunny
3:31
with a chance of being a good boy.
3:33
Here's the thing. No matter how many
3:35
steps we add, this is still just an AI
3:39
workflow. Even if there were hundreds or
3:41
thousands of steps, if a human is the
3:44
decision maker, there is no AI agent
3:47
involvement. Pro tip: retrieval
3:49
augmented generation or rag is a fancy
3:52
term that's thrown around a lot. In
3:54
simple terms, rag is a process that
3:56
helps AI models look things up before
3:58
they answer, like accessing my calendar
4:00
or the weather service. Essentially, Rag
4:03
is just a type of AI workflow. By the
4:06
way, I have a free AI toolkit that cuts
4:07
through the noise and helps you master
4:09
essential AI tools and workflows. I'll
4:10
leave a link to that down below. Here's
4:12
a real world example. Following Helena
4:14
Louu's amazing tutorial, I created a
4:17
simple AI workflow using make.com. Here
4:19
you can see that first I'm using Google
4:21
Sheets to do something. Specifically,
4:23
I'm compiling links to news articles in
4:25
a Google sheet. And this is that Google
4:28
sheet. Second, I'm using Perplexity to
4:31
summarize those news articles. Then
4:34
using Claude and using a prompt that I
4:36
wrote, I'm asking Claude to draft a
4:38
LinkedIn and Instagram post. Finally, I
4:42
can schedule this to run automatically
4:44
every day at 8 a.m. As you can see, this
4:46
is an AI workflow because it follows a
4:49
predefined path set by me. Step one, you
4:52
do this. Step two, you do this. Step
4:55
three, you do this. And finally,
4:57
remember to run daily at 8 am. One last
4:59
thing, if I test this workflow and I
5:02
don't like the final output of the
5:05
LinkedIn post, for example, as you can
5:08
see right here, uh, it's not funny
5:10
enough and I'm naturally hilarious,
5:11
right? I'd have to manually go back and
5:16
rewrite the prompt for Claude. Okay? And
5:20
this trial and error iteration is
5:23
currently being done by me, a human. So
5:25
keep that in mind moving forward. All
5:27
right, level three, AI agents.
5:29
Continuing the make.com example, let's
5:31
break down what I've been doing so far
5:33
as the human decision maker. With the
5:36
goal of creating social media posts
5:37
based off of news articles, I need to do
5:39
two things. First, reason or think about
5:43
the best approach. I need to first
5:44
compile the news articles, then
5:46
summarize them, then write the final
5:48
posts. Second, take action using tools.
5:51
I need to find and link to those news
5:53
articles in Google Sheets. Use
5:55
Perplexity for real-time summarization
5:58
and then claw for copyrightiting. So,
6:00
and this is the most important sentence
6:01
in this entire video. The one massive
6:04
change that has to happen in order for
6:06
this AI workflow to become an AI agent
6:09
is for me, the human decision maker, to
6:13
be replaced by an LLM. In other words,
6:16
the AI agent must reason. What's the
6:19
most efficient way to compile these news
6:20
articles? Should I copy and paste each
6:22
article into a word document? No, it's
6:24
probably easier to compile links to
6:26
those articles and then use another tool
6:28
to fetch the data. Yes, that makes more
6:30
sense. The AI agent must act, aka do
6:34
things via tools. Should I use Microsoft
6:37
Word to compile links? No. Inserting
6:39
links directly into rows is way more
6:41
efficient. What about Excel? M. So the
6:44
user has already connected their Google
6:45
account with make.com. So Google Sheets
6:47
is a better option. Pro tip. Because of
6:49
this, the most common configuration for
6:51
AI agents is the react framework. All AI
6:55
agents must reason and act. So
6:59
react. Sounds simple once we break it
7:01
down, right? A third key trait of AI
7:03
agents is their ability to iterate.
7:06
Remember when I had to manually rewrite
7:08
the prompt to make the LinkedIn post
7:10
funnier? I, the human, probably need to
7:13
repeat this iterative process a few
7:15
times to get something I'm happy with,
7:17
right? An AI agent will be able to do
7:19
the same thing autonomously. In our
7:22
example, the AI agent would autonomously
7:25
add in another LM to critique its own
7:28
output. Okay, I've drafted V1 of a
7:30
LinkedIn post. How do I make sure it's
7:32
good? Oh, I know. I'll add another step
7:34
where an LM will critique the post based
7:36
on LinkedIn best practices. And let's
7:38
repeat this until the best practices
7:40
criteria are all met. And after a few
7:42
cycles of that, we have the final
7:45
output. That was a hypothetical example.
7:47
So let's move on to a real world AI
7:50
agent example. Andrew is a preeeminent
7:53
figure in AI and he created this demo
7:55
website that illustrates how an AI agent
7:58
works. I'll link the full video down
8:00
below, but when I search for a keyword
8:02
like skier, enter the AI vision agent in
8:07
the background is first reasoning what a
8:10
skier looks like. A person on skis going
8:12
really fast in snow, for example, right?
8:14
I'm not sure. And then it's acting by
8:18
looking at clips in video footage,
8:22
trying to identify what it thinks a
8:24
skier is, indexing that clip, and then
8:29
returning that clip to us. Although this
8:32
might not feel impressive, remember that
8:34
an AI agent did all that instead of a
8:36
human reviewing the footage beforehand,
8:39
manually identifying the skier, and
8:42
adding tags like skier, mountain, ski,
8:45
snow. The programming is obviously a lot
8:47
more technical and complicated than what
8:49
we see in the front end, but that's the
8:51
point of this demo, right? The average
8:53
user like myself wants a simple app that
8:56
just works without me having to
8:58
understand what's going on in the back
9:00
end. Speaking of examples, I'm also
9:02
building my very own basic AI agent
9:05
using Nan. So, let me know in the
9:07
comments what type of AI agent you'd
9:08
like me to make a tutorial on next. To
9:11
wrap up, here's a simplified
9:12
visualization of the three levels we
9:14
covered today. Level one, we provide an
9:17
input and the LM responds with an
9:19
output. Easy. Level two, for AI
9:22
workflows, we provide an input and tell
9:24
the LM to follow a predefined path that
9:27
may involve in retrieving information
9:29
from external tools. The key trait here
9:31
is that the human programs a path for LM
9:34
to follow. Level three, the AI agent
9:37
receives a goal and the LM performs
9:39
reasoning to determine how best to
9:41
achieve the goal, takes action using
9:44
tools to produce an interim result,
9:46
observes that interim result, and
9:48
decides whether iterations are required,
9:51
and produces a final output that
9:53
achieves the initial goal. The key trait
9:56
here is that the LLM is a decision maker
9:58
in the workflow. If you found this
10:00
helpful, you might want to learn how to
10:02
build a prompts database in Notion. See
10:04
you on the next video. In the
10:05
meantime, have a great one.
