Andrej Karpathy floated a small idea that cleared 16 million views: stop using AI to write code and use it to build a second brain instead. Point Claude Code at a folder, drop in any source — an article, a transcript, a PDF — and let the model read it, link it, and file it into a living wiki of what you know. The pitch is that it compounds like interest. The more you feed it, the more it's worth.
The setup is real and it is trivial. You install Obsidian, create a vault, and open it in Claude Code. You paste the wiki instructions and tell Claude to build the scaffolding: a raw folder for sources, a wiki folder for the pages it writes, and a CLAUDE.md that defines how ingestion works. Then you drop a file into raw, say "ingest this," and ask questions across everything you've ever added. Five minutes, and you never start from a blank chat again.
I've run a version of this for months. The idea is correct. The framing — "I genuinely don't understand why everyone isn't using this yet" — is the part I'd push back on, because it locates the difficulty in exactly the wrong place.
The setup is five minutes; the system is months
The scaffolding was never the bottleneck. Personal knowledge bases have existed for decades, and the reason most people's notes rot isn't that creating folders was too hard. There are three real costs, and the viral thread skips all of them.
The first is the discipline tax. A second brain only compounds if you feed it on a schedule you actually keep. The "ingest this" step is frictionless precisely once; the hundredth time, after a long day, is when systems like this quietly die. Karpathy's framing makes the easy 5% look like the whole job.
The second is trust. An LLM that links and files your notes is also an LLM that can invent a backlink to a page that doesn't exist, compress a PDF into a confident summary that's subtly wrong, or file a source under a heading that misrepresents it. With code you get a compiler and tests to catch the model. With a knowledge base you get nothing — the error just sits there looking authoritative until you trip over it six months later. A second brain with no provenance trail is a liability that grows with use.
The third is context. "Ask across everything, forever" collides with the context window the moment "everything" gets large. Either the tool is doing retrieval over your vault (in which case quality depends entirely on retrieval you can't see) or it's stuffing files into the prompt (in which case "everything" is a marketing word, not a literal one).
What this gets right
Credit where it's due, because the core instinct is good. Your knowledge lives in plain markdown files you own, not in a closed app's database. The approach is model-agnostic — nothing here is locked to one vendor. Obsidian's backlinks give the "living wiki" claim a real substrate instead of a metaphor. And the compounding is genuine if, and only if, you do the unglamorous part. This is a better default use of an agent than yet another throwaway code-gen chat.
What builders actually said
Here's the uncomfortable part. The reaction to this idea, in the discussion I could see, is almost entirely vibes. The loudest reply is pure adrenaline: "You guys don't understand how life changing this is to me oh my God this is sick in both ways." [1] That post cleared 17,000-plus likes, and it reports a feeling, not a result. Nobody in the viral spread is showing me a vault that's six months deep, where the AI-written links held up and the cross-source questions returned something they couldn't have found by hand.
“You guys don't understand how life changing this is to me oh my God this is sick in both ways”
The pushback, such as it is, isn't technical either. One reply just shrugged at the evangelizing tone: "Not everyone has to understand you. You have nothing to prove." [2] That's the level of the conversation right now — sentiment, both directions, and almost no field reports. For an idea pitched on the strength of compounding over time, the absence of anyone actually reporting from month six is the most telling thing in the thread.
So I'll state the gap plainly: the evidence that this works long-term is currently zero, because the idea went viral faster than anyone could have tested it. What would settle it is boring and specific — someone showing a months-old vault, the bad links the model made, and a question that retrieval answered correctly across sources. Until then, "why isn't everyone using this" answers itself. People have been burned by note systems before.
If you're setting this up this weekend
Do it — it's a better use of Claude Code than most — but go in with eyes open:
- Treat every AI-written link and summary as a draft. Spot-check what it files for the first few weeks before you trust it.
- Keep the original sources in
rawuntouched, so you always have ground truth to diff the wiki against. - Decide your feeding cadence now. A second brain you update twice and abandon is just a folder.
- Watch how it behaves once the vault is large. That's where "ask across everything" gets tested, and where the real verdict lives.
The idea deserves the 16 million views. The certainty doesn't, yet.
Sources
- Tweet by @Poopiewoopief — @Poopiewoopief · community
- Tweet by @_iamhabeeb — @_iamhabeeb · community
- Tweet by @tanlovestannies — @tanlovestannies · community