Andrej Karpathy

The question worth asking about local models isn't whether they're as smart as the big ones. It's which of my 1000 calls ever needed the smart one. Andrej Karpathy puts it plainly: people are renting frontier models for… さらに表示

Andrej Karpathy Local Language Model Engineering 2026

The real shift isn't just about asking LLMs for answers anymore; it's about demanding they build the actual thing. Andrej Karpathy laid out a framework where you stop accepting plain text as the default and start… さらに表示

The graph in that Google paper is what got me. Qwen-9B with evolved skills posts 47.4% across five benchmarks. Qwen-27B running bare posts 39.4%. Both Qwen, neither one fine-tuned. Smaller model wins. What skill… さらに表示

Google WikiSkill paper SKILL.md agents

Andrej Karpathy put 10 years of experience at OpenAI and Tesla into one free lecture. people pay $15k for bootcamps that cover half of what's in it. you probably don't have 2 hours right now. don't let it disappear… さらに表示

Andrej Karpathy ChatGPT graph engineering

OpenAI users just realized they ignored Karpathy's free lecture for a whole year lol

Andrej Karpathy earned his PhD at Stanford working with Fei-Fei Li. He co-created CS231n, a course that taught an entire generation deep learning. As a founding member of OpenAI, he later departed to lead autopilot AI at… さらに表示

’s knowledge graph concept has been realized as Graphify, a fully open-source utility that integrates directly with Claude Code and Cursor. The workflow is simple: point the tool at any directory and it automatically… さらに表示

Graphify open source tool

Stop obsessing over GPT vs Claude vs Kimi. The real battle isn't about bigger context windows or slightly better prompts. Andrej Karpathy just exposed why your agents keep failing, and it has nothing to do with the model… さらに表示

Skip the Netflix episode and dive into Karpathy’s one-hour Stanford lecture on AI engineering. It is a solid weekend bookmark. He breaks down the reality: an LLM only delivers 10%, which is merely the starting point… さらに表示

Karpathy Stanford AI engineering lecture

My whole reading workflow is basically Claude + one Obsidian folder. Articles, PDFs, transcripts, all of it lands in that single folder. Claude pulls out atomic notes, hooks them into everything else, and then I just… さらに表示

Andrej Karpathy’s system relies on Obsidian to hold the vault while Claude reads and links every source you add. The setup is minimal: three folders and one file. Month one feels stagnant. You drop in a few articles, but… さらに表示

Karpathy breaks down how to evolve an LLM from simple Q&A into something that actually does stuff for you. The five ingredients are Memory, Multimodal, Reasoning, Search, and Tools, the full stack for going from… さらに表示

You’ve crushed 200 articles this year and can barely recall 5. Not a memory problem, those pieces had no place to land. Karpathy’s second-brain setup catches everything. Drop an article, PDF, or transcript into a single… さらに表示

Claude ripped through six years of my notes in a single night. Articles, abandoned ideas, book summaries I forgot I’d written, it surfaced connections I never would’ve caught on my own. The whole thing runs on Karpathy’s… さらに表示

A second brain isn't a galaxy, it's a living network of ideas and notes built by one person. Your brain takes in millions of bits per second but consciously holds almost none of that data. This system catches the… さらに表示