本文由 AI 分析生成
建立時間： 2026-05-25 來源： https://x.com/JoshKale/status/2042644740513149238
Summary
A summary of Karpathy’s “tractable brain upload” concept: collect all personal knowledge artifacts (articles, papers, repos, notes) into a folder, have an LLM compile them into a living wiki with inter-linked articles and a master index, then query the wiki rather than relying on the LLM’s training knowledge. The system self-audits for contradictions, fills gaps, and suggests new questions. Karpathy’s own implementation produced 100 articles, 400,000 words on a single research topic — all AI-written.
Karpathy「可行的大腦上傳」概念摘要：將所有個人知識收集到一個資料夾，由 LLM 編譯成有相互連結文章與主索引的活維基，然後查詢維基而非依賴模型的訓練知識。系統自我審計矛盾、填補空白並建議新問題。
Key Points
- 4-step process: collect all raw material → LLM compiles into linked wiki → query the wiki for specific questions → LLM audits and updates the wiki continuously
- Key property: agent doesn’t start from scratch each time — it has a pre-compiled map of the knowledge domain
- Karpathy’s version: 100 articles, 400,000 words on a single research topic, human typed none of it
- Karpathy’s framing: “lossy, approximate version of brain upload — tractable form available today via LLM simulators, without exotic neuroscience”
- Future implication: when you’re gone, your children could inherit an interactive map of your entire mind — every passion, fascination, and intellectual rabbit hole
Insights
This is the same concept as Farzapedia (@FarzaTV) implemented as a personal knowledge base, and as this vault’s own note-synthesis approach. The “brain upload” framing is philosophically interesting but the practical value is more immediate: a well-compiled personal wiki dramatically improves agent performance on personalized tasks by providing structured, traversable context instead of unstructured document dumps. The “audits itself” step (finding contradictions, filling gaps, suggesting new questions) is the compounding mechanism that separates a static archive from a living knowledge base.
Connections
Raw Excerpt
Decades of dreaming about brain uploads as science fiction and it turns out we just needed markdown files and an LLM that never forgets.