Andrej Karpathy figured out how to clone your brain.
The setup takes an afternoon and improves itself every day.
Here’s how:
Step 1: Collect everything. Every article you read, every paper, every repo, every note. Dump it all into a single folder. No organizing. No tagging. Just raw material.
Step 2: Point an LLM at that folder. It reads everything and “compiles” it into a living wiki. It writes articles about every concept it finds, creates links between related ideas, and builds a master index of everything you care about. You don’t write a single word. The AI is your full-time librarian.
Step 3: Ask it anything. The AI doesn’t start from scratch every time like ChatGPT normally does. It already has a map of the knowledge. And every question you ask makes the wiki smarter. Your curiosity literally compounds.
Step 4: The AI audits itself. It runs health checks across the entire wiki. Finds contradictions. Fills gaps. Suggests new questions you haven’t thought to ask yet. The knowledge base grows while you sleep.
Karpathy’s version on a single research topic: 100 articles. 400,000 words. He typed none of it.
We’ve been imagining brain uploading as exotic neuroscience decades away. But a lossy, approximate version that captures what you know, how you think, what you obsess over?
That’s possible today. Feed an LLM the full history of your thinking. Fine-tune it on your patterns. Now there’s an API endpoint that reasons like you and knows what you know.
Someone pointed out the most interesting implication: one day when you’re gone, your kids could inherit an interactive map of your entire mind. Every passion. Every fascination. Every rabbit hole you ever went down. A living conversation with the way you thought.
Decades of dreaming about brain uploads as science fiction and it turns out we just needed markdown files and an LLM that never forgets.
Andrej Karpathy @karpathy · 2026-04-10
Yes it’s the tractable form of brain upload. There’s a ton of scifi on brain uploads that requires way too exotic tech (scanning and simulating brains etc), when we’re about to get a lossy and approximate version of that *a lot* sooner via LLM simulators. You can easily imagine a
