AI for Building a Second Brain Webinar June 11th

This was a shorter, definitions-first session: around twenty minutes on what a second brain actually is, a live look at the one I built, and the three levels you can build it in. It is for anyone whose knowledge is scattered across notes, emails, transcripts, and a CRM, and who keeps hearing the term "second brain" without a precise definition. My usual toolkit applies here: ChatGPT, Perplexity, Gemini, and Claude on the everyday side, with Antigravity, Claude Code, Apify, Nano Banana Pro, and Codex on the advanced side.

What a second brain really is (1:34)

Lately a lot of people have been talking about second brains. "I built a second brain that does this, I used this model to create a second brain," and so on. What I wanted to do in this session is make sure you understand what a second brain truly is, and the fact that it does not matter which AI you use to create it.

Because a second brain is all about the knowledge and zero about the AI.

So the plan was simple: define what a second brain is, show you the one I created, and tell you the use cases it is actually useful for. Keep in mind this is evolving very quickly, because it is a brand-new branch of AI in quotes. But I already see quite a few use cases for this kind of architecture.

The LLM Wiki, Andrej Karpathy's approach (2:58)

If you have been following AI for the last couple of months, you will have heard the term LLM Wiki. The LLM Wiki is the first type of second brain, created by Andrej Karpathy, one of the founding members of OpenAI.

What Karpathy did, probably with Claude Code at the time, was take all of his pieces of paper. Pieces of paper meaning emails, sales transcripts, pieces of knowledge. He told the AI to read document one, then document two, then document three. By thinking very hard, as it usually does, the AI came up with a list of topics, tagged each document with its topics, and then created links between documents according to the topics they share.

That is the very foundation of an LLM Wiki and of a second brain. It has nothing to do with AI. It has everything to do with the simple fact that you have a lot of information stored in Google Drive, CRMs, and other places that is already linked, except you do not know about it. What AI does is help you find those links.

Building the index of connections (5:37)

Once those connections are established, everything gets consolidated into one file called the index. This is the first topic, this is the second topic, these are the links between documents, and so on.

When a new document comes in, the AI reads it, tags it accordingly, notices which existing topics it touches, and creates the links between the new document and everything related. Little by little you build up all the necessary connections between all of your knowledge, and this is what a second brain is. It is nothing more than that.

It is literally connections between pieces of knowledge that were already there before AI even existed. All the notes you ever took in Apple Notes, all your emails, all your transcripts, everything you saved on YouTube. AI only builds the connections between those pieces of knowledge, and you can plug any AI you want into it.

Live walkthrough of my second brain (7:03)

To demo mine, I asked Codex to open my LLM Wiki Visualizer. And yes, I still use both Claude Code and Codex. Quite honestly, the best value for money is still Codex: it gives you better usage, and it is more than fine for most business use cases. You do not need a PhD-level cybersecurity guy on your team if the only thing you want to do is analyze marketing campaigns. That is what I call killing a fly with a cannon.

What you see in the visualizer is nothing more than text files with a huge index. The index is connected to literally everything; it is the center of it all. Once you have created your second brain, you can tell Claude Code: before you answer anything, go check the index of my second brain and find the knowledge we already have.

The added value of a second brain is being able to talk with your business, in quotes. It is like having a digital double that embodies your business or your clients, and asking it active questions. Who is this person? What is the status of the pipeline? How is my content performing? And because everything is connected, you can link an answer back to related things like your content research.

Mine was built over several levels. Level one is the same LLM Wiki I just described, and then there are levels two and three. The skill I am sharing works for all three.

Aside: design systems and getdesign.md (8:39)

While Codex was opening the visualizer, I took a question from the chat: "I struggle to set up a good design system. I feed it brand guidelines, the front-end code base, screenshots, and the design is still off on colour, size and spacing. How do I fix it?"

My recommendation is a website that is part of my resources: getdesign.md. It explains in extreme detail what a design system is. What I did on my end was download all of those design systems and use them as inspiration for new ones. You have a lot of design systems that work, so you give one to Claude and say: this design system is perfect, it works. Now take the new assets I gave you and change this design system, or create a new one with it.

One more housekeeping note from the chat: my webinars are not public on YouTube, but they are always public on my website, mainly because private data can appear here and there in a live session.

The three levels. Level 1: keyword links (12:00)

Back in Figma, I walked through the three levels of the AI brain. I had run this explanation at a company a day or two earlier and was not happy with it, so this time the goal was to nail it so that even non-technical people can follow.

Level one is extremely simple. Say you have three pieces of paper. Paper one contains the word revenue. Paper two contains the word revenue. Paper three contains the word ARR. Level one is the Karpathy level, the LLM Wiki: the AI looks at every document and finds the words that are repeated across multiple documents, then writes the index with the corresponding links. Those links are done by word.

So it finds the link from revenue to revenue. The problem? ARR is clearly about revenue too, but we skip it. That is the limitation of level one. Even so, having just level one is already extremely useful.

A related comment from the chat: the index seems useful for RAG and setting up agents. My honest take is that no one uses RAG anymore. It is very rare to see a RAG architecture now, because we have things that just work better and are a lot simpler. AI is getting so good that you do not need to do a lot of complex stuff. The index is just useful to make sure your latest data is always available.

Level 2: embeddings, meaning instead of words (14:44)

Level two is where the mathematics kick in, and it is something I actually created. I am still questioning my own creation, but it does come up from time to time.

The way AI understands words is not like us. It does not know what the word revenue is; it does not even know what the word hi is. What it understands are two things: tokens and embeddings. A token is the building block, the Lego, of words for AI. The word love might be split into something like "l" and "ove". That might not be exactly accurate, but it gets you the idea.

An embedding is different. Think of it as a dictionary of words for the AI, a dictionary of what each token means. And the way this dictionary works is that it puts tokens with similar meanings close to each other. If you picture the dictionary as a literal two-dimensional room and you place the word cat somewhere, then dog will sit nearby, and lion too. A cat and a dog are different things, but both are animals with four legs and two ears, and a cat and a lion are both felines.

Here is why that matters for your second brain. If one document had a misspelling of the word revenue, level one would be blind to it. But in the embedding space those two words are so close that the AI understands they are the same. It also understands that ARR is similar, because it refers to revenue.

So level two gives the AI an embedding and has it check every document again, looking for words that are close inside that dictionary. This is brand new, I built it myself, I have been running tests, and it works well. It helps you establish new connections, but it is not there yet.

Level 3: document sectioning (19:05)

Level three is even simpler. Take a recording with your boss. You might start talking about project A, then the conversation drifts to theme B, and later to theme C. For level three, another AI reads every document and does this sectioning, so we understand that a single document can contain many topics. That helps us create even more connections inside the brain.

Those are the three levels. Quite honestly, if you start simply with level one, you will already be able to create and visualize the shape of your second brain. The skill I created for this visualization goes up on my website the day after the session.

Resources, my website, and the API pricing dashboard (20:30)

Before wrapping, I gave a quick tour of my website. On the Resources page you can go into AI Resources, which I am building out as I go, currently with Fable, making sure I cram the hell out of those tokens.

One example is the API pricing dashboard, which shows the evolution of pricing for every model out there. The new contender, Fable 5, comes in at a total cost per request of about fifteen dollars. The output is a hundred K, which is why the number looks like that, but as you can see it is by far not the cheapest model out there.

I will also add a dedicated section for the second brain use case, where you can pick up the skill, throw it into your Claude or your Codex, and use it. Beyond that, the site has all of my YouTube videos, searchable, and all the webinar recordings organized by theme: AI for business, automation, design systems, and so on. The webinar signup flow is also getting a new funnel, shape still to be decided.

Q&A highlights (23:07)

Do you still need RAG? Before, people used RAG to chunk documents and feed them into the AI. Now models have a context window so huge and such powerful visual understanding that this is literally built in. If you want to test it, take a five-hundred-page financial report that is available online, give it to Claude, and ask precise questions about it. It will understand. So yes, in most cases you can just put everything into the prompt.

What about agents that should only use internal information, not general search? For that, look into the Files API for Gemini and for Anthropic. They are built around creating internal knowledge bases. But quite frankly, the models are already good enough that if you connect them to the right tools for internal data and give them a nice system prompt, they will find the information very well and very fast.

Wrapping up (25:23)

That was the session. Shorter than usual, but the core idea is worth repeating: a second brain is one hundred percent about your knowledge and zero percent about the AI. Start with level one, look at the shape of what you already know, and build up from there.

If you want more sessions like this one, all the recordings are organized by theme on the webinars page.