YT YouTube2Text

YouTube2Text → Transcripts → Do You Actually Need AI Agents? (Why I Use Folders Instead)

Do You Actually Need AI Agents? (Why I Use Folders Instead) — Transcript

by Jake Van Clief · 7,421 words · 1,031 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00When I put this into an AI, there's a
0:03lot of different things that happen.
0:04Now, if you've been following me for a
0:06while, you know I use a lot of folder
0:08architecture. So, I'm actually going to
0:09start at that level uh of kind of, hey,
0:12I've designed a whole bunch of
0:13instructions into a folder, and my AI is
0:17going to follow those instructions every
0:19time I ask a question. So, in this case,
0:22let's say I have a folder, and it has
0:24instructions on how I like to build
0:27things, right? And that's just a folder
0:29with some markdown files, maybe a couple
0:31text documents that literally is just me
0:33describing this is how I like slide
0:35decks, maybe there's some assets in
0:37there, right, for different images,
0:40things like that. And that's a nice
0:41subfolder there. And I describe how I
0:43like to build things in it. Then I have
0:45another folder that describes me, right?
0:48And we'll have that be maybe my brand
0:51voice, where I'm from, my past, things
0:55like that. And again, just text files,
0:57just all organized neatly into an easy
1:00to read, navigable file. But then I
1:03might have certain artifacts I like to
1:06build with and write with. So a slide
1:10deck might have a skill, which if you go
1:13online and use Claude or Codeex or
1:16Gemini, there's things called skills
1:18that you can download that do certain
1:20tasks or create certain things. And
1:22there is one that you can get right now
1:24that says slide deck builder. And some
1:26of them usually build HTML, so big old
1:29code pages, and then convert that into
1:32PDF or convert it into a PowerPoint. But
1:36let's just say that's someone else's
1:38little folder process because if you
1:39open them up, it's just marked down in
1:41folders that says how to build a slide
1:44deck specifically. And the reason we had
1:46to describe that is because if I had
1:48asked this a year ago or two years ago,
1:51it has no idea what a slide deck looks
1:53like or if it does, it's kind of doesn't
1:55know the best way to put it, the best
1:57practices. Someone figured that out, put
1:59it into a box, which again is just a
2:02folder, and then gave that to you,
2:04right? Okay. And if you've watched my
2:05other videos, we describe how you need
2:07data, context, and tools to make an AI
2:11not stupid. And in other words, to make
2:14it actually create good stuff. This was
2:16one of the solutions to do it. So,
2:18right, we've gotten our build folder,
2:19maybe a subfolder on how to build
2:21slides. We can call that a skill. Maybe
2:23you built it yourself. Then again, my my
2:27is a little different, right? If I said,
2:29"Build me a slide deck for my investor
2:31into an AI that had none of this folder
2:34structure, none of this knowledge."
2:36Guess what? Doesn't know who you are.
2:38Might know how to build a slide deck.
2:40Okay? Definitely doesn't know who your
2:42investor is or your company. Again, you
2:44can swap these words out. But if I have
2:47a nice little folder structure, then
2:49guess what? I can simply describe what
2:51that is. So, let's go ahead and expand
2:53me out again. Folder inside a subfolder.
2:58much simpler here. So under me, I might
3:01have something about the things that I
3:03work with, the people that I work with.
3:05And in this case, I have a folder about
3:09my investors for my company. Or you
3:12could swap it out for clients. But
3:14either way, I have maybe a couple
3:17markdown files, text files that describe
3:19who they are. Maybe there's links to
3:20their websites just pasted in there. All
3:23of that sort of stuff. And so that
3:25folder or subfolder is part of this kind
3:28of process, my company, the way I like
3:31to work things. And every time I ask my
3:33AI, whether it's Claude or Anthropic or
3:36anything, put it straight into the chat.
3:38As long as it has access to this folder,
3:40it will go through and go build me.
3:43Okay. What do you like to build and who
3:45are you? A slide deck. Okay. How do I
3:48build slide decks for my investors?
3:52Okay. my you have investors let's read
3:55who those are traditionally over the
3:58last two or three years this was a much
4:00more complicated process now
4:03traditionally if I wanted my AI model
4:08to be able to do a task like this we
4:11would have to create some sort of agent
4:13or at least that's what a lot of people
4:15have done so they would take the AI and
4:18the model on itself knows nothing about
4:20you maybe sucks at building slide decks
4:22definitely doesn't know anything about
4:23your investors unless it has to go do a
4:25bunch of research. So over the last two
4:27or three years, what people have done is
4:30they would create a database that maybe
4:32has a whole bunch of information on you.
4:35And this could be an SQL database, which
4:38for my data nerds out there, you know,
4:40it's just a really efficient way to use
4:42code to capture data and bring it
4:44towards you. They might use a big giant
4:47Google Drive that has a whole bunch of
4:49stuff. maybe a SharePoint or any sort of
4:53other kind of data place. Other times we
4:56put it into these text files called JSON
4:58and YAML. We don't need to go into that,
5:00but just know it's a little bit more
5:03complex and structured than those
5:05markdown files or text files that we
5:07were just talking about. problem is
5:09these are all behind an API or some sort
5:12of basically gateway that says, "Hey, in
5:15order to access me, in order to get to
5:17me, you have to say these words in this
5:19order to make sure you get the right
5:21stuff." That goes for all of these
5:22databases. So, the AI doesn't know how
5:25to do that. So, we got to create a
5:26little package for it that lets it do
5:29that. And usually, we would write that
5:31in code. So, this could be Python, C, a
5:34computer language. Uh, and I've written
5:37quite a lot of these, right? I used to
5:39do this all the time. I have a master's
5:40thesis which I, you know, I have a
5:42couple videos on. You can go check them
5:43out. Uh, where I took a whole bunch of
5:45different psychometric scales, you know,
5:47personality tests and essentially
5:49created a bunch of pipelines to access
5:52all those tests and send them to the AI
5:55model. Now, again, the reason we had to
5:58do this is because the AI doesn't have
6:01this information trained into it. You
6:04can train a good bit of it, but even
6:06then it's still not as efficient.
6:07Sometimes it messes things up if you
6:09purely just train it on data. So
6:11sometimes it's more efficient and better
6:13to give it to it. If we don't give it
6:15that data, it doesn't know anything
6:17about me. The next problem that we have
6:19to deal with is building a slide deck.
6:22So it has the information on me. So now
6:26we're done here. We built this really
6:27cool Python connector. We've connected
6:29it with code. It it's able if I ask,
6:32hey, who am I? It's able to pull that
6:34information. We solved that problem. Now
6:36I need it to be able to build a slide
6:38deck. So then I create another bit of
6:40code that says this is how to build a
6:42slide deck. Here's how it connects to
6:44Microsoft Power King. Let's use a red
6:47one there. So we'll do uh I don't know
6:50if PP is the right word here, but we're
6:52going to go for it anyway. So we've got
6:54a PowerPoint and you could do Google
6:55Slides. There's a lot of steps. There's
6:58assets. There's clickable buttons.
7:00There's so many different things. This
7:01is it using another software. And so at
7:03first we tried to get the AI to control
7:06making PowerPoints. I worked with a
7:08large company that uh Fortune 500. We
7:11have an NDA so we can't talk about it
7:12too much, but I trained over 1500 of
7:14their employees over the course of a
7:16year. And we were using Microsoft
7:18Copilot which had a pipeline data
7:21connection and a very specific way to
7:23manage PowerPoints. The worst
7:25PowerPoints I have seen an AI make ever.
7:28They were awful because it couldn't
7:30fully control the whole system. It's
7:32gotten significantly better now. But the
7:34problem is it can't interact with
7:38PowerPoint software. Well, so to make
7:40that better, we have to create a
7:42separate tool that allows it to either
7:46write code that gets converted into a
7:48PowerPoint or take pictures that are
7:51pasted into a PowerPoint. And so common
7:54things were to write HTML code, the
7:56stuff that websites are written with, or
7:59to just build it out in some sort of
8:02HTML or image model, take a screenshot,
8:05and then feed that and paste it into a
8:07PowerPoint because the AI just could not
8:10really work with PowerPoint. Well, it
8:11just it's just it didn't work. We didn't
8:13have the systems. This with enough
8:16iterations now, some of them do work
8:18pretty well. They can decently mess with
8:20PowerPoint in that way. and they have a
8:22decent amount of tools, but really all
8:24you needed was a couple of these tools.
8:26And I'll say why that's important in a
8:28second. But again, we're still not done.
8:30It had all the information on me, but
8:32what about these investors? Well,
8:33there's websites, there's uh maybe
8:36different LinkedIn profiles. They might
8:37have all of this stuff. So, we need to
8:39give it a tool that lets it scrape the
8:42web. And so, that might be more code,
8:44more Python, more C, whatever it is. And
8:47it rolls out and then it finally finds
8:50that investor. hopefully sometimes it
8:52didn't. So then we had to have error
8:54catching built into it to make sure that
8:56it would find the right person based on
8:57my question. Also my investors needs to
9:01go back over here to pull and double
9:02check that it's actually calling them.
9:05And so that's what it would do and it go
9:07out and it would find it. Finally, once
9:09it's taken all of that information in,
9:11it consumes it all and then it says,
9:14"Hey,
9:16here's who your investors are. Here's
9:18the slide deck. Here's what you've
9:19made." Now, I don't know if I need to
9:21tell you this. That's one agent. If I
9:24change this word to a PDF document, oh,
9:28I don't have a tool for PDFs. So, I have
9:30to come back over and build a whole new
9:32one for how that works. The instructions
9:34for how I like to do it, connect it to
9:36these ones, make sure it's reconnected.
9:38So, now I have agent one, which builds
9:41side decks and knows about my investors
9:43and handles all these things, and agent
9:45two that can build PDFs in the same way.
9:47Now, I will say once you built the
9:49architecture for the first one, it's
9:51easier to build subsequent agents, but
9:54why would we do that when instead we can
9:57flip it all on its head? So, let's go
9:59back to the folder again and why I think
10:03it's such a powerful tool in the AI
10:06world. Between 1969 and 1970, you have
10:10two men named Ken Thompson and Dennis
10:12Richie who discovered a really good
10:14philosophy called Unix. And basically
10:16they said if you can separate data, put
10:19it into a nice file tree structure and
10:23put it onto a page that you can click
10:25through, you have a great way to
10:27interact with a computer. And they
10:29created all the structure, all the
10:31ideas, thought about how you would
10:32interact it with your mouse, how it
10:34would work on operating systems and
10:36Windows because again you might take for
10:38granted folders on your laptop didn't
10:41always exist. Folders weren't always an
10:43option for people, right? They haven't
10:44always existed and we take for granted
10:46some of the most simple software that we
10:48use, but it's often the most powerful
10:50because we take it for granted. I mean,
10:52think about it. If you didn't have
10:54folders, wouldn't your entire workflow
10:56collapse? I know most of these
10:57multi-million dollar companies, it goes
10:59straight down the drain real quickly. If
11:02that's the case, if something invented
11:04in the ' 70s is still fundamental today,
11:07just because an AI is reading the text,
11:10navigating it, using the software,
11:12doesn't mean that fundamental
11:14disappears. It just means something else
11:16is using it. So to be able to do that, I
11:19did some testing and I said, well,
11:21instead of making all this crazy agentic
11:23stuff, what are the minimum tools that
11:25an AI needs to be able to read a folder?
11:28And I'm not the only one who's doing
11:30this. Google is doing this, Anthropic is
11:32doing this. OpenAI is diving deeper into
11:34it a lot now because they kind of have
11:36to because of how many people and
11:37researchers are exploring it. But the
11:39idea was we can build a harness. So a
11:43harness is simply a set of general tools
11:48that seem to be used almost over every
11:51workflow. So, a tool that lets you read
11:54folder files, a tool that lets you write
11:59into folder files,
12:01maybe one tool that lets you search the
12:04web or search online. And those are
12:08essentially a package. And no matter
12:10what AI you use, where you use it, all
12:13of the best ones at least have this
12:15little harness attached to the model.
12:18Most AI that we talk about today is just
12:21a big model with a whole bunch of
12:23numbers that when you put something in,
12:25it decides the best next step to give
12:27back to you, right? And it does that
12:29thousands of times. So when I say,
12:31"Build me a PDF for my investors, as
12:34long as it has a couple of these tools,
12:36I don't need to build a whole bunch of
12:37agents." It'll go through and read,
12:39write, and search for what it needs. And
12:42if you can give it what it needs before
12:44it has to go look elsewhere, guess what?
12:47you don't need to build all this crazy
12:48stuff. So, let me show you how that
12:50would work. And so, in order to show you
12:53the massive interesting side of this,
12:55let's try to condense what the AI model
12:57is so that you can appreciate why giving
13:00it a harness does so much. So, you have
13:03in 2017 a paper called attention is all
13:07you need. This came out from Google
13:10Labs, Google DeepMind I believe. And
13:12essentially without diving into the huge
13:15amount of math sitting inside this
13:17paper, there has been decades of AI
13:20research. I mean all the way back in the
13:231950s when we first coined artificial
13:25intelligence. All the way to the 1980s
13:29and 1990s where we started talking about
13:32neural networks. copied what we thought
13:34neural pathways in the brain would look
13:36like as code to modern-day transformers,
13:39which is what was talked about in that
13:412017 paper. They basically said, "Hey,
13:44taking this decades and decades of
13:46research on how AI should work, what
13:50predicting the next process should be,
13:51how can I use language to control a
13:54computer?" They said, "Well, we can just
13:56take all of that data. We can take every
13:59time you make a decision, every time you
14:02decide something and we can give it a
14:05number and that number can be a
14:08percentage of its possibility happening
14:10again. So if I was to say the
14:14[snorts] sky
14:16is
14:18blank, I can have a list of words based
14:22off thousands of sentences that have
14:24ever been said and have a percentage.
14:26Obviously, the highest percentage is
14:29probably going to be blue. Let's say
14:31that's 60%. The next highest percentage
14:34might be beautiful
14:37or gray
14:40or any of those. And so each of these is
14:43going to have a different percentage of
14:45possibility when I ask a question or say
14:48a statement. And so what we did is we
14:50took all of those up and then we said
14:53that's one set of numbers. But now, what
14:55if I added a sentence in front of this?
14:58The UK sky or the night sky. Context
15:06changes the percentage. So, let's go put
15:09a whole bunch of numbers around all of
15:11that. And so they did that billions and
15:14billions of times to get every possible
15:17combination and context layer and said,
15:21"Okay, well, if we can do that enough
15:23and we can get enough numbers every time
15:26you put in a question, the model should
15:28output the most likely answer, which
15:32more often than not is the correct
15:34answer." The thing is, the more we put
15:37layers on it, the more we did that, the
15:40more it gave not the most likely answer,
15:43but the most seemingly fitting answer,
15:46which there's a difference there. When I
15:49believe something is correct, I'm
15:52thinking based on what I have learned.
15:54So if you add one more layer to this,
15:56which is the person using it,
16:00you know, a lot more context
16:02information, which means the right
16:04answer predictively, the UK sky is gray
16:09most often. But if they know it's me and
16:11I'm studying in spring in Edinburgh and
16:14it was a beautiful day and I talk about
16:16how much it is, then my most likely
16:18answer is going to be beautiful. The
16:20right most likely answer is not too much
16:24different from how we think, how we look
16:26at data. The hard part is creating the
16:30massive, massive databases and the
16:33algorithms that look through all of
16:34that. That is years and years of
16:38research, years of effort, years of
16:41failure. And yes, sometimes they suck,
16:44sometimes they are wrong. But to sit
16:46there and diminish it as simply a word
16:49predictor is to diminish the years of
16:52effort and research. To diminish it down
16:54to that is to diminish the years of
16:56research that people and effort have put
16:59into it. The time and the love that we
17:02have looked at because we're all looking
17:04at engineering meaning structure. Which
17:08brings us back to the original thing.
17:10Little side note there. When I make a
17:12folder to hold all of my context, who I
17:16am, what I'm doing, how I like to do
17:19things, I'm providing another layer of
17:22data to help it come to my right answer,
17:26what I think is best. And this is the
17:30opposite of what a lot of people are
17:31worried about with cognitive thought and
17:33the models controlling you. We can flip
17:36it on its head and we can decide what
17:39its context is. we can be the one in
17:41control and it becomes much more of a
17:43tool rather than a predictor. So why
17:47would I want to do that? Let's go back
17:49to the original folder structure. If I
17:51sit there and I have my instructions for
17:53building, I have my instructions for who
17:55I am and what I like. I have my
17:57instructions for how PDFs and things
17:59should be built.
18:01And then I have my harness
18:04that has the ability to read, right?
18:07That might actually be code, right? This
18:09is the one place where code is needed.
18:11But this is the same as downloading a
18:13software or tool. It's it's really not
18:15too hard nowadays. Most of those models
18:17and AI already have them built in to
18:20write
18:21and to search.
18:24Then technically with just these few
18:27things, as long as my instructions are
18:29there, it can do almost anything I need.
18:32So, if we're looking at opinions,
18:34instructions, and data, and we're just
18:36describing them in text files inside of
18:38a folder, when I say, "Build me a PDF
18:40for my investors," it goes into the
18:43folder and only reads the opinions
18:46associated with what I asked, what my
18:49instructions are, and what my data is.
18:51So, let's say I have hundreds of
18:53different questions and things that I
18:55like to do on a daily basis. Instead of
18:57having to build a whole bunch of crazy
19:00connectors and integrations, I can just
19:03add a folder and describe what I want to
19:06do. And that's it. I just added a new
19:08feature, a new agent essentially. As the
19:12model, as the AI goes through the
19:15folder, goes through the instructions,
19:18it's reading what it needs to and
19:21becomes
19:24your PDF agent, your slide deck agent,
19:27your uh marketing agent. Instead of
19:31having to build agents for every
19:32process, you're building instructions
19:34for each process and letting one agent,
19:37one AI become what it needs when you
19:41need it, when you want it to. Because in
19:44a world full of cheap answers, questions
19:47become valuable, instructions become
19:49valuable. This places the person at the
19:52center of decision-making. This is an
19:55augmentation rather than just an
19:58automation. I think we're pretty clear.
20:00It's much simpler to build out a folder
20:02to handle a lot of processes. Obviously,
20:04there's specific scenarios, but let's
20:07start following it down into the rest of
20:09the computer, right? We talked about the
20:11instructions. We talked about the
20:12folders and where it is, but let's talk
20:15about the other side of this. How is
20:17this being processed? What does that
20:19look like? What did we have to do
20:21beforehand to actually have this code?
20:23Now, we're not going to dive in too
20:25deep. I'm not going to get absurdly
20:26technical, but I think it's important to
20:28understand the thousands of lines of
20:30code that actually happen when you put
20:33this in. And so instead, we're going to
20:36replace this with traditional code. One
20:38of the most basic ones that everyone
20:40learns about, and I actually have
20:42another video on this where I go in more
20:43detail. Let's just talk about it real
20:45quick. Print hello world. This is one of
20:50the most well-known computer statements
20:52that you learn in almost every
20:54programming class. for this situation.
20:56This is written in Python. Now, Python
20:59was created in the 1990s. It was
21:01actually named after Monty Python, uh
21:03not the snake. Uh the creator of it
21:05really much enjoyed the movies and
21:06figured that would be a funny way to
21:08talk. And essentially, if we talk and I
21:11say print hello world, this is a
21:15command. This command says, "Hey
21:18computer, you need to display in an
21:21image uh below this the words hello and
21:25world." There's a lot going on here.
21:28Now, when you learn it in computer
21:30programming, all you see come out of the
21:33computer
21:35is hello world. Like, ah, it worked.
21:39This is actually thousands of lines of
21:41code that took 40 years to develop not
21:44in Python but in everything else behind
21:46it. In the same way if I asked an AI to
21:50write this code statement in the same
21:53way people are vibe coding now or asking
21:55it to write code when that code is ran
21:59even after the AI built it there's
22:01hundreds of things happening. So let's
22:03go ahead and look at what that looks
22:06like. So, Python is a computer language.
22:10Now, before we go into that print hello
22:13world, it's important to understand why
22:15someone would make Python and what a
22:18computer language is. So, let's go even
22:20farther back to 1800s.
22:24I know, so far away. There was a man
22:27called Jakard and he created something
22:30called the Jakard Loom. I believe it was
22:31somewhere in France. I can't remember
22:33the name off the top of my head right
22:34now. And essentially he said, "Hey,
22:37right now we've got all these people and
22:39they're using these crazy looms and
22:41they're going back and forth and they
22:43have to keep moving these levers back
22:44and forth to change the design. So, what
22:47if I could make a beautiful little card
22:53that had some holes in it that you could
22:55slot in and essentially all the springs
22:59on those little levers would go up or
23:01down whether or not there was a hole. So
23:04anywhere there wasn't a hole was flat,
23:07the lever would be up. And anywhere
23:09there was a hole, it would slot into the
23:12hole and go down, which on the other
23:13side meant it would lift up, changing
23:16the thread pattern. And if you did that
23:18enough times and you made certain areas
23:20flat, you would eventually have a card
23:24that would have a pattern on it up and
23:28down. Created an image, created a series
23:31of interactions. So, as I sat there and
23:33folded back and forth, the loom would
23:35roll over itself in a certain pattern.
23:38This is one of the first ever computer
23:40programs. Granted, no electricity, no
23:43programs technically, but it took logic.
23:47He had to understand where each hole
23:50would end up on the other side. He had
23:52to understand how you could translate
23:53something physical, an image or a
23:56pattern into a series of up and down.
24:00Fast forward a little bit, only a couple
24:03years, about 30 years, we had someone
24:06called Adah Love Lace, absolutely
24:10brilliant mathematician, and she worked
24:12on something called the analytical
24:14engine. Basically, it was this giant
24:17physical gear thing with a whole bunch
24:20of gears in it. And each gear had a
24:24little sprocket on it that would
24:26represent a number zero through nine.
24:29And so if you had one gear here, it
24:32would be zero, one, two, four, depending
24:36on where that gear is. And as you turned
24:38it, it would change them in an order.
24:40They would use some sort of card that
24:44had holes in it to decide the starting
24:47pattern of the numbers, and it could
24:49then do calculations. Of course, the
24:52amount of cards that you would have to
24:54make to change all of the numbers based
24:56on gearing was massive. And so she was
24:59tasked with writing an amazing book
25:01around it. Actually, specifically the
25:03notes because, well, they didn't really
25:05like women being authors of books at the
25:07time, unfortunately, despite she was
25:09absolutely brilliant. So instead, she
25:11wrote the notes of the book. The book
25:13was about 120 pages long. Her notes were
25:16470 pages long. So the real book was the
25:20actual notes. What's really cool is I
25:22did a research project on her. I read
25:24through a whole bunch of her notes. We
25:25saw patterns with her mathematics still
25:27being used in a lot of algorithms we
25:30look at today. But the coolest thing is
25:32every so often she'd be writing notes,
25:34all these mathematic notations. And the
25:35next thing you know, hm, you know, if
25:38you assigned the right variables in the
25:40right place, I bet you this could
25:42produce music. It could pluck a string
25:43in the right order. It could do an
25:45orchestra. And then she'd go right back
25:46into the math and she'd be writing and
25:48writing and writing and algorithms. And
25:50then she would stop again and she would
25:51return back to that thought and just
25:53write another note and say, "Well, does
25:54that mean if it can make music that
25:56they're thinking?" Well, what would it
25:58need to be to be a thinking machine? And
26:00that statement and what she argues about
26:03there actually goes back
26:07before the 1800s. This is important
26:10because if you can learn the history of
26:13the why, the how becomes much easier to
26:16understand. I could sit here and try to
26:19teach you how to program in Python, how
26:22to build an AI, how to run it, but if
26:24you don't understand why, how doesn't
26:27help you at all. If you don't understand
26:29the why, you don't understand how
26:32something breaks. You don't understand
26:33what it took to get there and why we do
26:36certain things certain ways. It also
26:38means you don't know what you can skip,
26:41what you don't need to learn because
26:43someone has already figured it out and
26:45it's just a process in the giant cog. So
26:48that idea of thinking machines that Ada
26:50Love Lace talked about actually was
26:53brought up for the first time in the
26:551640s, I believe 1642 by a man named
26:59Renee Deart. You know, I think therefore
27:02I am. You might have heard that before.
27:04Well, he didn't just write on
27:05philosophy, religion, and mathematics.
27:08He also wrote on thinking machines in a
27:11book called Discourse on the Method. He
27:14wrote about two possible ways to know if
27:17a machine was thinking. Mind you, he
27:20didn't test it. He was just watching a
27:22whole bunch of cool puppets and weird
27:24steam machines and these Da Vinci
27:25Renaissance engines going around and he
27:27said, "If it had a human limb, if it had
27:31a human face, what if what if we were to
27:33build all the organs that a human would
27:35have? How do I know it's not human? How
27:37do I know it's not thinking? And even if
27:38it's not human, what's the difference?"
27:40And so he said, albeit in 17 pages, much
27:44longer than I'm about to describe it
27:46here, well, first of all, it would have
27:48to, and I quote, not communicate like a
27:50brute. And what he meant by that is have
27:53a conversation, understand what I mean,
27:55not just respond to commands, but to add
27:58to them, to to be able to be a partner
28:01in thought. But he said, we still might
28:03be able to create that in a machine. He
28:04said, if we could organize every single
28:06word and made sure it organized the
28:08right way, then maybe we could do that.
28:09that. So, it needs to do a second thing.
28:11It needs to be able to do something
28:14outside of what we designed it to do or
28:16what it was designed to do. So, a clock
28:20is not thinking because it can tell the
28:22time, but being able to tell the time is
28:24a very powerful thing for thinking
28:26beings. We built the clock because we
28:29couldn't keep track of it. So if the
28:31machine can do both of those things,
28:33build something outside of itself and
28:36what it was built to do and uh could
28:38talk to us, well then you know what
28:40maybe it is thinking. Ada Love Lace
28:43going back to her and her lovely Note G
28:47is one of the most famous ones where she
28:49talks about this. If you look up
28:50analytical engine and Google Note G,
28:53you'll find a lot of information on it
28:54and you can go read it yourself or you
28:56can go check out I'll have some links in
28:57the bio. I made some papers on, so you
28:59can just jump straight to that. But she
29:01said, you know, it might be able to make
29:03music. It might be able to predict what
29:05it could be doing. It might be able to
29:07do all sorts of fantastical things, but
29:10it can't originate anything. It always
29:13needs the intention of me, of a human.
29:17And even if we built all of these
29:18machines, they were built with the
29:20intention of a human. So the spark of
29:22consciousness, the spark of light, no
29:24matter how advanced they get, was with
29:27us. So there is no origination.
29:31Funny enough, if you fast forward all
29:33the way to the 1940s,
29:36someone picked up on this long
29:38conversation, 1940s,
29:41maybe you've heard of him before.
29:43There's quite a few movies about him,
29:46but his name is Allan Turing. There is a
29:50specific test called the Turing test.
29:53Now he did not call it that. He called
29:55it the imitation game. But he said if
29:58you as a computer could talk to a person
30:02and on the other side of it you don't
30:04know if it's a human or a person, it's
30:07passed the consciousness test. It is the
30:09imitation of thought that is more
30:12important than the actual thought being
30:14done. Because he believed we still don't
30:16know what thought is. you could sit here
30:18and describe it and what it is. And he
30:21actually specifically references Adah
30:23Love Lace in one of his papers. I
30:25believe it's his 1952 paper where he
30:27says it's called the Ada Love Lace
30:29objection. He goes, you know, she didn't
30:31believe that they could originate
30:32anything. But there is nothing original
30:35under the sun. And that's a great Greek
30:38philosopher. All of us are listening to
30:40programs of school, of where we grew up,
30:42of what is being given to us by the
30:44environment. So what is original to
30:47begin with? He didn't attack saying that
30:50the AI couldn't originate. He said that
30:52there's no such thing as original, that
30:55all of this is imitations and building
30:57together and movements. It wasn't
30:59necessarily an argument that they could
31:01be consciousness. Rather, he said it is
31:03a common fallacy that both
31:06mathematicians and philosophers fall to
31:08that once all the facts of a system are
31:11known, all of the consequences are known
31:13with it. But that's not true. We often
31:16do research for decades. We find every
31:18little thing and we come to find out
31:21that in fact we are surprised. We missed
31:24something. Oh, we didn't get all the
31:26data. And his argument is that machines
31:28could surprise us too despite of what we
31:31know. But around this time there's
31:34another important person that came about
31:36and this is where we come into
31:37programming and processes. At this point
31:40machines come together. Well, we had
31:42decided that we could take the jakard
31:44loom and we could add zeros and ones
31:47together. We could instead of up and
31:48down, let's assign it to a number value.
31:51And then if we assign patterns of those
31:53zeros and ones to another number, we can
31:56start doing math with it. So we assigned
31:59something called binary. And so from 0
32:02to 15, we could use four indications off
32:06and on to create 15 numbers. That's a
32:09really cool logic. And if we use the
32:11algorithms and the processes of Ada Love
32:14Lace and Jakard and then we put it into
32:16electronics, switches and cables, well,
32:18we could start making the computers do
32:20math. It could do algorithms. It could
32:22it could do all of these things. And
32:24Alan Turing used this to break the Nazis
32:26enigma. He created a massive machine
32:28that could calculate its cryptography,
32:31all the secret message. And it was
32:33impossible to do that hand by hand
32:35because there were so many possible
32:37calculations. But if a machine did it,
32:39it could do it faster, but they still
32:42augmented touring. They didn't automate
32:45it. And so what ended up happening is
32:48saying, "How do we make that smaller?
32:50How do we make it more condensed?" Now,
32:52many of you might be using an AI called
32:54Claude. Well, that is actually named
32:56after a wonderful individual called
32:58Claude Shannon. Now, when we talk about
33:01information theory, when we talk about
33:03logic and ideas and switches, Claude is
33:06one of the pioneers in all of the major
33:09ways we still even use today. Claude
33:11talked about the ways in which we
33:13interact with semantics and meaning are
33:15completely separate from the engineering
33:17problem. If it can be described in
33:19logic, it can be made into hardware. And
33:23Claude decided to go ahead and write all
33:25the papers we would possibly need on
33:27that. Not all of them, but a really huge
33:29amount of them. Now, I don't know what
33:31was in the water in the 1940s or 50s,
33:33but I can tell you we had some heavy
33:35hitters. There's another one called
33:37Grace Hopper, an absolutely brilliant
33:41individual. But Hopper was dealing with
33:44a similar concept that I'm dealing with
33:46today. So, Grace Hopper came up with
33:48something called the compiler. And
33:50essentially what a compiler does is it
33:53says, "Hey, you have all of these words
33:55and commands. What if we routed that to
33:59all of that crazy machine logic you had
34:01before? Now, one of her first compilers
34:05was something called the Flowmatic. Now,
34:07essentially, you had all of these crazy
34:09instructions, right? These zeros, these
34:12ones, these ifs, these ends, these RBNs,
34:15these moves, blah blah blah. We had a
34:17whole list of things that we were doing
34:18in machine language, assembly language,
34:21things like that. And she said, 'Well,
34:23there's a lot of patterns. In the same
34:25way we saw patterns in zeros and ones
34:27and wrote machine code, she said, well,
34:29it doesn't just have to be arithmetic.
34:32We can assign it very specific
34:34sentences. We can say move or guide or
34:40place or insert small little sentences
34:44that would stand
34:47for a whole bunch of specific actions,
34:51specific code being ran. But in order to
34:54make sure that that all mapped well, in
34:56the same way we had to map those zeros,
34:59those ones, those ideas to the logic and
35:02write that all out, she had to write out
35:05all of the ways in which this would map
35:08to hundreds of code in different orders.
35:11And what you're starting to notice here
35:13hopefully is there's a similarity.
35:15Language to action, language to action,
35:20right? We're building on it here. And so
35:22she created one of the first real
35:24compilers, as in I will take all of your
35:27words and compile it down, take it all
35:30together and decide what actions the
35:33machine could take. And there's
35:35thousands of lines of code that would
35:36happen with that, right? Funny enough,
35:39there's a quote from her and she says,
35:41"I had a working compiler and no one
35:44would touch it. They said computers
35:46could only ever do arithmetic."
35:49Gracehopper had built the system, had
35:51seen how it worked, had gotten it
35:53working, and decided, you know, this
35:56should be something shared with other
35:58people. And the traditional engineers of
36:00the 30s, 20s, and 40s didn't believe
36:02her. They said, it's too unknown. You
36:06can't control the code. It's going to
36:08make too many errors. It's not in
36:10control of it anymore. You're not
36:12thinking about the code anymore. Does
36:15that sound familiar? I don't know. Maybe
36:17we'll find out. Well, long story short,
36:19it ended up being one of the most
36:21important inventions and led to all
36:23sorts of amazing languages. We got
36:26Forran, which is still being used by
36:27NASA today to be able to send ships up
36:30into space. We got all sorts of amazing
36:33computer languages that allowed us to
36:35write instructions in code that would
36:39handle all of that crazy binary and
36:42absurd things we were talking about
36:43earlier. Now, once we understood that we
36:46could compile something, as in we could
36:48take pieces of code, pieces of patterns,
36:50and it didn't all need to be written
36:52down all the time, we just needed to
36:54catch the patterns. Someone else could
36:56write it down and we could compile that
36:59into machine readable language as in
37:02those silicon transistor, those files,
37:05those processors go, "Oh, zeros and
37:07ones. I get that. It doesn't understand
37:09English." We started building a lot more
37:12and people realized well why are we you
37:14know these are still very mathematic
37:16languages even 4chan and all of these
37:18lower level languages were still really
37:20hard for people to use and learn it was
37:23more math than it was language so that's
37:26when languages like C and Python started
37:29coming about now C and it was adding a
37:33couple more English words that we could
37:36actually use to do things like the word
37:39print being something that is a command
37:42that controls a whole bunch of assembly
37:45and machine code to make the computer do
37:48something, make calculations, put an
37:50image on a screen and handle text. But
37:54it was still not there. It was still a
37:56little challenging. So our lovely friend
37:58I talked about earlier said, "Well,
37:59let's see if we can make it even more
38:01readable. Let's focus around objects."
38:03And we call it objectoriented
38:06language. as in we have a thing that
38:09does a certain stuff i.e. separations of
38:12concerns. Let's just put that behind a
38:14word. So let's take all of this C stuff
38:17and actually just make that into a word
38:20as well. So people were building a whole
38:22bunch of different programs that did
38:24really good stuff in C and they were a
38:26whole bunch of lines of code. Well, we
38:29keep reusing that program again. What if
38:32we made that a statement? made that a
38:35word, right? And so, Python kind of does
38:38that. It's not exactly like that. So you
38:40do want to take this with a grain of
38:42salt. But essentially when you write
38:45something in Python,
38:48it's actually taking that and it's
38:50deciding what else did you write? What
38:52was here? And it's making a tree. And
38:55it's deciding, okay, well based on this
38:58is a certain type of command or output
39:01or process. And it's assigning that a
39:03value. And it's just doing that for
39:05every single sentence you write. And
39:08then it's compiling that down into by
39:11bite code which is then getting compiled
39:13technically down into machine readable
39:15code which is then getting compiled down
39:17into the zeros and ones which is then
39:19just electrical signals moving back and
39:22forth which is really just off and on.
39:25So every single time you type a key,
39:30the letter K, the letter J into a
39:33keyboard, that is hundreds of lines of
39:36code being executed. You are living on
39:39the top of giants right now. Every
39:42process has been logically built out. We
39:45have been using that process all the way
39:48up into today. Constantly abstracting
39:50one idea and building on top of it and
39:53building on top of it and building on
39:56top of it on the ever journey to get
39:58closer. So that one thing, what is the
40:01goal of computer language? To turn our
40:04language into something a machine can
40:07do. That has always been its goal. That
40:11has always been its direction. We've
40:13tried to do that hundreds of times
40:14before and we keep doing it. But it also
40:18automates the processes of intent. Each
40:22stage was complicated. So we added that
40:25into a sentence into a word and brought
40:28that into a smaller version of the next
40:31stage which became more complicated. And
40:33then we condensed all of that down until
40:35finally we find ourselves here today. We
40:38are at one of the epidium of computer
40:41languages that people had wanted so long
40:43ago which is I can say build me a PDF
40:47for my investors and it will run all of
40:50this code that is a line of code now our
40:54English language has become code but
40:58there's one problem AI doesn't know what
41:03you think is best what you want your
41:06intention
41:08just like with all the other software is
41:10still the main goal. So how can we
41:13standardize giving my intention to this
41:17AI? My folder method that I explained
41:20all the way at the beginning of the
41:21video is one method in the history of
41:24time to standardize the way in which we
41:28structure our intent, our logic, our
41:31zeros and ones, all the symbols around
41:35it. Because when I say build, it's
41:38compiling down into a whole bunch of
41:40stuff. I'm simply describing that in
41:43folders. When I say PDF, I'm simply
41:47describing how to build a PDF. And it is
41:50connecting to tools that already exists.
41:53The best engineering in the world
41:56happening with AI is happening in
41:59English language. The reason computer
42:02language has been so powerful for us,
42:04the more people that can use it, the
42:06more things we can build to automate to
42:09give to people to augment. Our goal for
42:13today, our goal for the next decade is
42:16to learn how do my intentions and my
42:19instructions and AI fit
42:23into the next step. Because what happens
42:27when we turn this
42:29into a word called print? What does that
42:33look like? What is the next abstraction?
42:35And the only people who are going to
42:37figure that out is us because we're the
42:40ones standing on the back of giants. And
42:42in the future, people will be standing
42:44on our backs. And until next time, happy
42:48learning.

About this transcript

This page contains the full transcript of Do You Actually Need AI Agents? (Why I Use Folders Instead) by Jake Van Clief, generated from the public captions YouTube serves with the video. The transcript has 7,421 words across 1,031 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.