YT YouTube2Text

YouTube2Text → Transcripts → Costruisco un Second Brain Avanzato prima che scoppi la Bolla AI | EP. 2

Costruisco un Second Brain Avanzato prima che scoppi la Bolla AI | EP. 2 — Transcript

by Antonio Guadagno · 2,907 words · 478 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Today we are going to build a "second
0:03brain" together, a digital brain that
0:05can literally remember every piece of
0:07information about us.
0:10Who we are, what we do for a living,
0:12what projects we are working on, and so
0:14on.
0:15And the reason we are building this
0:18second brain is so that we can then
0:21connect it to our AI agents, so that
0:23when we ask one of these agents a
0:26question, it won't give us a generic,
0:28trivial response, but will instead query
0:31the second brain, find the information
0:34it needs, and give us personalized
0:36answers relevant to our context.
0:41In this video, therefore, I will guide
0:44you step by step in creating a second
0:47brain that we can use with any agent,
0:49Claude, Copilot, Codex, ChatGPT, and
0:52many others.
0:55Specifically, I will be using it with my
0:58Hermes Agent, and that's because this
1:01video is part of a series I'm doing on
1:04the channel where I build my own custom
1:07AI agent.
1:09This is the second episode.
1:11If you missed the first one, I'll leave
1:13it below in the description, and if you
1:14don't want to miss the next ones,
1:16subscribe to the channel.
1:18First of all, what is a second brain?
1:21In practical terms, it is a personal
1:24knowledge base that concerns us, our
1:26work, and our context, which is managed
1:29and consulted by artificial intelligence
1:33.
1:33The advantages offered by a second brain
1:35are, in fact, two.
1:37First, that the information is managed
1:40by AI, meaning we dump our raw files in
1:42there as they are, and it is the AI that
1:45cleans them, organizes them, links them
1:48together, and so on.
1:52The second advantage is that by working
1:55with clean, organized knowledge with the
1:58right connections, it becomes much
2:00easier to find the correct information
2:03and we are also much faster at finding
2:06what we need.
2:10Regarding the structure we will give our
2:13second brain, I personally like to use
2:15this model called LLM Wiki proposed by
2:18Andrej Karpathy, formerly of OpenAI and
2:21a brilliant mind in artificial
2:23intelligence.
2:25According to this model, our second
2:27brain is divided.
2:30into three folders: raw, wiki, and
2:33output.
2:35The "raw" folder is the entry point to
2:38our second brain, meaning we upload our
2:41raw files here just as they are—those
2:43files we want to clean and bring into
2:46our second brain.
2:49These files are cleaned and organized by
2:52artificial intelligence, not by us, and
2:55are moved into the "wiki," which
2:58contains the processed, organized, and
3:01structured knowledge.
3:04We can therefore say that the wiki is,
3:05for all intents and purposes, our second
3:07brain.
3:10Finally, we have the "output" folder,
3:12which will contain any answers to the
3:15questions we ask whenever those answers
3:18require the production of a report, a
3:20summary, a document, and so on.
3:24Getting into more detail, the structure
3:26of the wiki is composed as follows.
3:29We have a main folder with a file called
3:32index.
3:33MD. We will return to this shortly.
3:36Inside this folder, we have many
3:38subfolders, specifically one subfolder
3:42per topic.
3:44For example, we have the client
3:46subfolder which will then contain all
3:49documents regarding clients, or the
3:51meetings subfolder which will contain
3:53all documents regarding meetings.
3:58Besides subfolders, another key element
4:01of the wiki are the indexes, which act
4:03as a compass that the artificial
4:05intelligence uses to navigate within
4:08this second brain.
4:11Indeed, we have a main index under the
4:14wiki folder and then many indexes, one
4:16for each subfolder.
4:19The main index describes the list of all
4:22the wikis, meaning all the subfolders,
4:25and explains which types of documents
4:28are contained in each subfolder.
4:32As for the subfolder indexes, these
4:35explain which files are contained in
4:37that subfolder and what type of
4:40information these files contain.
4:45Essentially, when we ask our AI agent a
4:47question, it enters the second brain,
4:50does not read every single document, but
4:53instead reads the indexes, and by
4:55reading the indexes, it understands
4:57which subfolder to go to and then which
5:00files are relevant to effectively answer
5:03the question we asked.
5:06With the concept understood, let's get
5:08our hands dirty and create this second
5:10brain.
5:12To create the second brain, I will use
5:14Herms Agent within VS Code.
5:19Actually, VS Code isn't really necessary
5:21; I only use it because, at least here
5:23on the left, we can see the files,
5:25folders, and so on, and therefore this
5:27video becomes easier to understand and
5:29follow.
5:31But you have the possibility to do
5:33everything directly from the Claude
5:36application, ChatGPT, or whatever agent
5:39you prefer. The first request I make is
5:42this one.
5:44Simply with this prompt, I ask the
5:46artificial intelligence to create the
5:49tree structure of my second brain,
5:51specifically the raw, wiki, and output
5:54folders and other files needed for the
5:56second brain.
6:00together.
6:02You can find this prompt below in the
6:04description; it is absolutely free, and
6:06you will also find the link to my
6:08Telegram group and my WhatsApp group,
6:10which are free as well.
6:12So, what do I do?
6:14I copy and paste this prompt, go back to
6:16Herms, paste it, press enter, and wait a
6:19few seconds.
6:21And so here is the structure of our
6:23second brain.
6:25We have raw, wiki, and output, and then
6:28we have two very, very important files
6:32called agents.md and log.md. The
6:34agents.md file will contain the
6:37instructions needed for Herms Agent,
6:40Claude, Codex, and so on, to understand
6:43how to manage and consult the second
6:46brain.
6:48And I will give you these instructions
6:50now.
6:51As for the log.md file.
6:56Inside this file, the artificial
6:59intelligence will write down every
7:01change it makes to the second brain, so
7:04that if necessary, we can read this file
7:07and understand what has changed.
7:11The instructions to put inside agents.md
7:14are these, and you can find them for
7:16free in the description below.
7:19First, we explain to the agent what its
7:21role is.
7:22It will be the librarian that manages
7:24this knowledge base.
7:26We then go on to explain how this
7:28knowledge base is structured.
7:31We explain that there is the raw folder
7:33containing raw files, the wiki
7:35containing the second brain articles,
7:37and the output folder to be used for
7:39responses.
7:41We then go into detail on how the wiki
7:44folder works with the indices, the
7:46structure of the articles, the links
7:48between them, and so on.
7:52And then, scrolling to the bottom, we
7:54also have the description of how the log
7:56works and the description of the three
7:58main actions you can perform with the
8:00second brain:
8:04compilation, consultation, and audit.
8:07As for compilation, it allows you to
8:10take files from the raw folder, clean
8:12them, organize them, and then move them
8:15into the wiki.
8:18As for consultation, it is used to
8:20search for information within the second
8:22brain, first by analyzing the indices,
8:25then analyzing the files, and so on.
8:29And finally, we have the audit or lint
8:32command, which is used to clean up our
8:35second brain, meaning that sometimes we
8:38want to check for inconsistencies, data
8:40that isn't coherent, files that might be
8:43old, therefore no longer current or
8:46valid, and so on.
8:49At the bottom, as I mentioned, there is
8:52also the description of how the log.md
8:55file works and therefore how it must be
8:57managed and how events that modify the
9:00second brain must be stored.
9:04What we do with this large prompt, very
9:07simply, is paste it into agents.md.
9:12We then save this file with command S,
9:15and at this point, we have the structure
9:18of our second brain and have written the
9:21operating rules.
9:23The only thing left to do is to run some
9:26tests.
9:27We will then load some files into the
9:30raw folder to compile them and see how
9:33they are handled.
9:36The most attentive and tech-savvy among
9:38you will have noticed that I am not
9:40using Obsidian, which for many is
9:42synonymous with Second Brain.
9:46The fact is that that tool is not
9:48functional for building a second brain,
9:50but at most, it serves to visualize
9:52files and see that nice graph with the
9:55files and connections.
9:57In practical terms, however, it is not
10:00essential for what we are doing right
10:02now. Currently, therefore, I don't feel
10:04the need for it, so I won't use it.
10:08If in the future I want to view my
10:10second brain files more easily, then I
10:13will add it by selecting it as a
10:15voltrain.
10:18To test this second brain, I uploaded
10:22three files to the row folder containing
10:26made-up transcripts of a phantom client
10:29that doesn't actually exist, as they are
10:33transcripts I had the artificial
10:36intelligence create. What do I want to
10:39do?
10:40I want to take these transcripts and
10:43move them into the wiki, organizing and
10:46structuring them in a more defined way
10:48that is easier to consult.
10:53So, I go to chat with Hermes and write:
10:56"If you look inside row, you will find
10:59three files
11:04containing call transcripts with the
11:07client Meridiana
11:13Assicurazioni."
11:16I want you to take these three
11:20transcripts, process them with the
11:24compile command, and move them into the
11:28wiki in a folder dedicated to the client
11:33Meridiana Assicurazioni.
11:38I press enter and I want you to note the
11:41level of detail I used for this prompt.
11:45I didn't just tell it to run the compile
11:48command and see what it could come up
11:51with; knowing how I work and that I want
11:53a wiki for every client, I am the one
11:56who imposed my way of working and
11:58managing the second brain on the AI.
12:02I told it how to create the wiki.
12:05In my opinion, this isn't a flaw, but
12:08simply a way of ensuring that artificial
12:11intelligence adapts to our processes and
12:14the way we work.
12:16And so, Hermes Agent has finished its
12:19work.
12:20We have the wiki folder containing the
12:23Meridiana Assicurazioni wiki.
12:26We have one file for assessment, one for
12:28training, and one for governance.
12:32We obviously have an index.md that
12:34explains the content of each file, and
12:37the interesting thing is that for each
12:40file we see the creation date, update
12:43date, we see the sources of this file,
12:46and then, of course, we see the content
12:49, which has been cleaned and structured
12:52ad hoc.
12:55We also have an update to the wiki index
12:58, explaining what the Meridiana
13:00Assicurazioni wiki contains, such as
13:03information on the client's AI adoption
13:06project with assessment, pilot design,
13:08and so on.
13:12And obviously, inside the log file, we
13:14have the operation I performed, which is
13:17this one, with all the information on
13:19the changes that were made, which are
13:22quite a few.
13:25Essentially, therefore, our second brain
13:28is ready, and what I will do now is
13:30delete these example documents and
13:32instead populate it with my personal
13:34documents, which I obviously won't show
13:37you because they are private.
13:40At this point, we have our second brain
13:43.
13:45As long as we're working with 10, 100,
13:47or 1,000 files, it works perfectly, but
13:49then, at a certain point, its
13:51performance worsens.
13:55Because, yes, it's true that we have
13:57indices that act a bit like a compass,
14:00but to find information, our agent still
14:03has to read these documents; and so,
14:05when the files we work with start to
14:08become too many, the performance of the
14:10second brain drops drastically, and it's
14:13better to use solutions like, for
14:15example, RAG, or Retrieval-Augmented
14:18Generation.
14:21To solve this problem a bit, the CEO of
14:24Shopify came up with a solution called
14:27QMD.
14:29For those who don't know what QMD is,
14:32the acronym stands for Query Markdown
14:34Documents.
14:36Essentially, it is a semantic search
14:39system, a keyword search system.
14:42within the markdown documents that make
14:45up our second brain.
14:48To put it simply, it's as if we were
14:51unlocking a Ctrl + F, a keyword search,
14:54but one powered by artificial
14:56intelligence.
14:59So, to find a certain paragraph or a
15:01specific topic that interests us, we
15:04don't need to type the exact words from
15:06that paragraph.
15:09Just a few keywords are enough, and our
15:11artificial intelligence will be able to
15:14find that information accurately without
15:17needing to read the entire document from
15:20top to bottom.
15:22Basically, it's as if we combined the
15:25second brain with a RAG system, allowing
15:27us to find precise information while
15:30consuming fewer tokens.
15:33And it is precisely this QMD system that
15:36we are now going to implement in our
15:39second brain, and the prompt we are
15:41using to set up this QMD system is this
15:44one, which, like all the others, is free
15:47below in the description.
15:50The first thing we do is specify which
15:53folder needs to be indexed, which is the
15:56wiki.
15:57We ask Hermes to install QMD.
16:02We ask to be asked for confirmation
16:05before each step, and we specify that we
16:08also want to install the QMD MCP, so
16:11that we don't have to remember the
16:14commands for specific wiki searches, but
16:18simply ask Hermes, and he will run those
16:21commands.
16:25Once we have copied this prompt, we take
16:27it to Hermes, I paste it, hit enter, and
16:30wait for the artificial intelligence to
16:32complete its work.
16:35So, QMD is now installed, and at this
16:37point, all that's left is to test it.
16:41What did I do?
16:43I updated my second brain by uploading
16:45the slides from my AI Evolution course,
16:47which is my practical course on AI
16:49applied to work.
16:51If you'd like, you can find information
16:52below in the description.
16:54I uploaded these slides in raw format,
16:57had them added to the wiki, and then,
17:01using QMD, I asked it to index this wiki
17:04.
17:05How do we know if it worked?
17:08Quite simply, if we go to the terminal
17:11and I run QMD status, it tells me that
17:14it has indexed the wiki and indexed six
17:17files, which are the four lessons plus
17:20the two index files.
17:24If I clear the screen here and run QMD
17:27search iteration,
17:33it actually searches for this keyword
17:36within my wiki, sort of like a Ctrl + F
17:40boosted by embedding and therefore by AI
17:43.
17:45Obviously, however, we don't want to use
17:48these commands, we don't want to
17:50memorize them. So what do we do? We use
17:52the MCP that has been installed on
17:55Airmess Agent.
17:57So here I can ask, use QMD and search
18:02for information on iteration within the
18:08wiki.
18:11It will then do what I previously did
18:14manually via the terminal; it will take
18:16care of running those commands itself.
18:20The big advantage, therefore, is that we
18:23perform a more precise search by keyword
18:25or at least by words similar to the ones
18:28we remember.
18:31We save a lot of tokens and when
18:33necessary, we can also manage to pick
18:36out the specific words we need.
18:40And so, Hermes is responding.
18:44It effectively reports the passages it
18:46read within our wiki.
18:50The only slightly annoying thing about
18:54QMD is that when we modify the wiki,
18:57perhaps by adding files, the wiki must
19:00be re-indexed, so we come back here into
19:03Airmess Agent, specify that the wiki has
19:07been updated, and ask to re-index it
19:10through QMD.
19:13To wrap things up, I would like to
19:15answer a couple of comments that a lot
19:17of people made under the first episode
19:20of this miniseries.
19:23And the first comment was: "Antonio,
19:25sorry, but what's the point of setting
19:27up this agent, this whole setup and
19:29everything, if in the end you use Open
19:31Router APIs?"
19:33And the answer is that it actually
19:35doesn't make sense, but that was only
19:37the first episode.
19:39I am currently building the system, I am
19:42setting up the infrastructure.
19:44I am doing it with OpenAI APIs because
19:47they allow me to test things quickly and
19:50very cheaply, but then, moving forward
19:53in the series, what I would like to do
19:55is buy a Mac Studio or, in any case,
19:58some powerful hardware to host this
20:01whole system.
20:02Locally with local models, so either
20:05Llama or LM Studio. Obviously, though,
20:09that is the end goal; it cannot be the
20:11starting point, and so there will be
20:13episodes that lead us to that point.
20:17Also, before I buy a Mac Studio, which
20:19by the way costs a ton of money, I’d
20:20like to get a sense of whether you’re
20:22interested in this series and if it’s
20:24helpful to you in some way.
20:27So, if you find it helpful and you're
20:29interested, leave a like and let me know
20:31down in the comments below.
20:33As for the second comment, this is
20:36actually a piece of advice that I truly
20:39appreciated a lot.
20:41What you told me, in fact, is "Antonio,
20:43rather than connecting to Gmail, Drive,
20:46and so on using the Google Cloud Console
20:49, which is a very cumbersome process,
20:52use Composio instead."
20:55This way you connect to Composio, you
20:57simply ask Hermes in the chat, and it
20:59handles the connection for you, and then
21:02through Composio you connect to
21:04practically all your applications via
21:06single sign-on, so basically with one
21:08click.
21:11I did it, and I have to confess that
21:13connecting to Gmail, Drive, and Calendar
21:16was much more convenient and much faster
21:19, so it's a tip I really appreciated.
21:23If you have any other tips, whether on
21:25the first or the second point, leave
21:27them down here in the comments.
21:30That said, I’ll sign off and see you in
21:32the next one.

About this transcript

This page contains the full transcript of Costruisco un Second Brain Avanzato prima che scoppi la Bolla AI | EP. 2 by Antonio Guadagno, generated from the public captions YouTube serves with the video. The transcript has 2,907 words across 478 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 choose a caption track available on the video.

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, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.