Andrej Karpathy Stopped Using AI to Write Code — He's Using It to Build a Second Brain Instead (2026) - AI Money Making

By - ziqingbo
Posted on 29/04/2026
Posted in AI Tools

Andrej Karpathy Stopped Using AI to Write Code — He’s Using It to Build a Second Brain Instead (2026)

## Table of Contents
1. [The Shift That Surprised Everyone](#1-the-shift-that-surprised-everyone)
2. [What Is a “Second Brain” in 2026?](#2-what-is-a-second-brain-in-2026)
3. [Karpathy’s New Workflow: No Vector Databases, No RAG Pipelines](#3-karpathys-new-workflow-no-vector-databases-no-rag-pipelines)
4. [The Markdown + LLM Stack That Actually Works](#4-the-markdown–llm-stack-that-actually-works)
5. [Why This Matters More Than AI Coding Tools](#5-why-this-matters-more-than-ai-coding-tools)
6. [How to Build Your Own AI Second Brain in 2026](#6-how-to-build-your-own-ai-second-brain-in-2026)
7. [The Best Tools for Building Your Second Brain](#7-the-best-tools-for-building-your-second-brain)
8. [Who Is This Actually For?](#8-who-is-this-actually-for)
9. [The Honest Downsides](#9-the-honest-downsides)
10. [Conclusion: Stop Using AI Wrong](#10-conclusion-stop-using-ai-wrong)

—

When one of the most respected AI researchers in the world decides to change how he uses AI — people listen.

Andrej Karpathy, former Tesla Autopilot lead and longtime AI educator, recently revealed that he’s fundamentally shifted how he uses large language models. He’s stopped treating AI as a code-writing tool and started using it as a thinking partner — a system for building what he calls a “second brain.”

This isn’t just a productivity tip. It’s a signal about where AI is actually heading in 2026.

## 1. The Shift That Surprised Everyone

Karpathy’s transformation didn’t happen overnight. For years, he was the guy who showed the world how to use AI to write better code — his tutorials on neural networks, LLMs, and computer vision have millions of views. He was the poster child for “use AI to code faster.”

Then he changed his mind.

“I stopped using AI to write code,” he explained in a recent discussion. “Now I use it to think.”

This might sound counterintuitive, especially coming from someone who literally taught the world about AI. But when you look at what he’s actually doing, it makes perfect sense.

**The insight**: Most people use AI as a faster version of Stack Overflow. Karpathy is using it as an extension of his own cognition.

## 2. What Is a “Second Brain” in 2026?

A second brain is exactly what it sounds like — an external system that stores, organizes, and helps you retrieve information and ideas. Think of it as a digital extension of your memory and thinking capacity.

But in 2026, the definition has evolved significantly:

The key difference is that an AI second brain doesn’t just store information — it helps you *think* with it.

## 3. Karpathy’s New Workflow: No Vector Databases, No RAG Pipelines

Here’s what makes Karpathy’s approach different from what most “AI second brain” products try to sell you: he deliberately avoids the complex infrastructure.

No vector databases. No retrieval-augmented generation (RAG) pipelines. No custom embeddings. No expensive vector storage costs.

Instead, he uses something far simpler: **plain markdown files and an LLM**.

This is the critical insight that most AI productivity influencers miss. They talk about building sophisticated systems with Pinecone, Chroma, or Weaviate. They build RAG pipelines with LangChain. They spend weeks setting up infrastructure.

Karpathy’s answer: **You don’t need any of that.**

His system consists of:
– A folder full of markdown notes
– A well-structured LLM (typically GPT-4o or Claude)
– A simple prompt that tells the LLM to reason about his notes

That’s it.

## 4. The Markdown + LLM Stack That Actually Works

The beauty of Karpathy’s approach is its simplicity. Here’s the actual stack:

### The Storage Layer
– **Markdown files** stored in a logical folder structure
– Files organized by topic, project, or date
– Each file contains a mix of:
– Meeting notes
– Research snippets
– Ideas and observations
– Code snippets
– Reading notes

### The Intelligence Layer
– **GPT-4o** or **Claude** (Sonnet or Opus)
– A system prompt that defines the AI’s role as “reasoning over my notes”
– No special tools or plugins — just the raw LLM

### The Interaction Pattern
Instead of searching for a specific note, you tell the AI what you’re thinking about. It reads through your notes and reasons about them — finding connections you might have missed, summarizing relevant points, and helping you think through problems.

**Example workflow**:
1. You have a question about AI safety research
2. You paste your question + relevant notes into the LLM
3. The LLM reasons over your specific notes and provides context-aware insights
4. You iterate and refine

This is fundamentally different from Google search or even Perplexity because it reasons over *your specific knowledge and observations*, not the general internet.

## 5. Why This Matters More Than AI Coding Tools

Here’s the uncomfortable truth that most “AI productivity” content ignores:

**AI coding tools make you faster at executing ideas you’ve already had.**

AI thinking tools help you have better ideas in the first place.

This is why Karpathy’s shift matters. When you use AI to write code, you’re treating it as a labor-saving device. When you use AI to think, you’re treating it as a cognitive amplifier.

The difference is profound:

And in 2026, with LLMs getting significantly better at reasoning, the leverage point has shifted from “AI does the work” to “AI helps me think.”

## 6. How to Build Your Own AI Second Brain in 2026

Ready to build your own system? Here’s a practical, no-nonsense guide:

### Step 1: Start with a Note-Taking System You Actually Use

Don’t overthink this. The best note-taking system is the one you’ll actually use consistently.

Options that work well:
– **Obsidian** (local, flexible, free)
– **Notion** (collaborative, clean UI)
– **Apple Notes** (simple, fast)
– **Plain text files** (universal, never lock in)

Karpathy uses markdown files specifically because they’re universal and never lock you into a proprietary system.

### Step 2: Develop a Consistent Note-Taking Habit

Your second brain is only as good as the information you put into it. This means:

– **Daily notes**: At the end of each day, write 3-5 bullet points about what you learned, observed, or wondered about
– **Meeting notes**: After any significant meeting, capture the key takeaways
– **Reading notes**: When you read something interesting, write a brief summary + your reactions
– **Idea capture**: When you have an idea, write it down immediately — even if it’s half-formed

The goal isn’t perfect organization. It’s capturing raw material for future thinking.

### Step 3: Configure Your LLM

You don’t need any special tools. Just set up a conversation with your LLM using a system prompt like:

“`
You are my thinking partner. I will share my notes and questions with you.
Your job is to help me reason through problems, find connections between ideas,
and think more clearly. Be direct. Challenge my assumptions. Don’t just agree with me.
“`

### Step 4: Start Reasoning Over Your Notes

When you have a problem to solve or a decision to make:

1. Open your notes folder
2. Find relevant notes (or dump your entire folder if it’s small enough)
3. Paste them into your LLM along with your question
4. Ask the LLM to reason about your specific situation
5. Iterate and refine

### Step 5: Build This Into Your Daily Workflow

The magic happens when this becomes a daily habit, not a one-time experiment:

– **Morning**: Review yesterday’s notes while your memory is fresh
– **Weekly**: Do a “thinking session” where you reason over a week’s worth of notes
– **Before big decisions**: Use your notes + AI to think through implications

## 7. The Best Tools for Building Your Second Brain

While Karpathy keeps it minimal, here are the tools that make this approach more practical:

### For Note-Taking

### For AI Reasoning

### For Syncing (Optional)

If you want cross-device access:
– **Obsidian Sync** ($8/mo)
– **iCloud** (for Apple Notes users)
– **Git** (for technical users — markdown files + git = free forever)

## 8. Who Is This Actually For?

Let’s be honest: this approach isn’t for everyone.

**This works great for:**
– Knowledge workers who consume a lot of information
– Researchers or anyone who needs to track complex topics over time
– Entrepreneurs who need to make sense of partial information
– Writers, strategists, and thinkers who want to improve their ideas
– Anyone who feels overwhelmed by information but doesn’t want to miss important insights

**This probably isn’t for you if:**
– You don’t regularly take notes
– You prefer to think in your head rather than write
– You need real-time information retrieval (this is for deep thinking, not quick lookups)
– You work primarily with structured data rather than ideas and observations

## 9. The Honest Downsides

No system is perfect. Here’s what you should know before diving in:

### It’s Not Search
This is the biggest misconception. Your second brain doesn’t give you instant answers like Google. It helps you *think through* complex problems. If you want a quick fact, use Google. If you want to reason about your situation, use your second brain.

### Consistency Is Hard
Building and maintaining a note-taking habit is genuinely difficult. Most people start strong and then stop after a few weeks. The system only works if you actually use it.

### Context Windows Are Still Limited
Even with 200K+ context windows, you can’t dump your entire life’s work into an LLM at once. You need to be thoughtful about what you include in each session.

### Quality of Input = Quality of Output
If your notes are messy, vague, or inconsistent, your AI reasoning sessions will be too. The system amplifies whatever you put into it.

### It’s Not Magic
Despite what some influencers claim, building a second brain won’t make you a genius overnight. It’s a tool that helps you think more clearly — you still have to do the actual thinking.

## 10. Conclusion: Stop Using AI Wrong

The mainstream narrative around AI productivity in 2026 is still largely focused on “use AI to do things faster.” Write code faster. Draft emails faster. Create content faster.

Karpathy’s shift reveals a deeper opportunity: **use AI to think better.**

This is a fundamentally different relationship with AI. Instead of treating it as a tool that executes your commands, you’re treating it as a partner that helps you reason through problems.

The best part? You don’t need any special infrastructure. You don’t need to pay for expensive vector databases or build RAG pipelines. You just need:

1. A note-taking system you’ll actually use
2. An LLM with good reasoning capabilities
3. The discipline to capture your thoughts consistently
4. The willingness to actually think, not just delegate

That’s the real lesson from Karpathy’s pivot. The most powerful AI tool isn’t the latest coding assistant or content generator. It’s the one that helps you think more clearly about the problems that actually matter.

**Stop using AI to do more. Start using AI to think better.**

—

*What’s your experience with AI-powered thinking tools? Have you tried building a second brain? Share your insights in the comments — I’d love to hear what’s working (and what isn’t) for you in 2026.*

**Related Articles:**
– [Build Your First AI Agent in 2026: A No-Code Step-by-Step Guide](https://yyyl.me/archives/3613.html)
– [The Complete Guide to AI Agents in 2026: From Zero to Full Automation](https://yyyl.me/archives/)
– [How Multi-Agent Systems Are Replacing Single AI Tools](https://yyyl.me/archives/3608.html)

AI Money Making - Tech Entrepreneur Blog

Andrej Karpathy Stopped Using AI to Write Code — He’s Using It to Build a Second Brain Instead (2026)

Previous Article

Next Article

Leave a Reply Cancel reply

news

archive