Tag
Karpathy used Claude to analyze 43 million Hacker News comments from the past 20 years, identify 6,454 repeatedly recommended videos, and rebuild HackerTube into a better YouTube recommendation tool in about 2 hours.
The tweet shares Andrej Karpathy’s two-hour hands-on AI course, which systematically explains the core workings of AI agents, harnesses, loops, and graphs. It is well suited for developers looking to gain a deeper understanding of AI agents.
The tweet shared a complete AI degree curriculum that Andrej Karpathy made freely available on GitHub, consisting of 7 repositories arranged in learning order, covering everything from backpropagation to agents capable of conducting research autonomously, with over 300,000 stars combined.
A long-form Chinese opinion piece drawing on Andrej Karpathy's tips for interacting with large language models (such as using the ASD-STE100 simplified language standard) to explain how AI can be used to clarify concepts, visualize them, and turn them into interactive mini-tools — enabling truly personalized learning tailored to each individual.
Andrej Karpathy built an agent called autoresearch that runs AI research autonomously using just 3 files — you give Claude or Codex a small LLM training setup and the agent loops through try, commit, and train.
A tweet highlights Andrej Karpathy's claim that prompting is becoming obsolete in favor of agent graphs, noting he built a graph from scratch in 69 minutes to show it as the emerging structure for agentic systems.
A practitioner reports that Karpathy's ASD-STE100 controlled-language tip genuinely improves LLM response clarity (including HTML artifacts) when applied selectively, and shares a concrete self-improvement prompt that samples recent transcripts to derive custom clarity rules for the user's AGENTS.md file.
A tweet from @cyrilXBT praises a paper on graph engineering for agentic systems, claiming it significantly improves AI responses, and promotes a spammy bidding website called RankNow.
Andrej Karpathy built an LLM from scratch and open-sourced all code, covering the full pipeline of dataset, training loop, and inference loop, continuing his work on simplifying AI education with micrograd and makemore.
Andrej Karpathy released nanochat, a minimalist LLM training framework that can train a GPT-2-level model for just around $48, covering the entire pipeline of pretraining, fine-tuning, and reinforcement learning, with minimal and fully transparent code.
A compilation of Andrej Karpathy's recent tweets covering a minimal GPT implementation, vibe coding an iOS app, and a detailed comparison of GPT-4.5 with earlier models, offering insights into AI testing and scaling.
Andrej Karpathy experiments with Opus 5, giving it the first paragraph of Lord of the Rings and a 1M token budget to create a 3D JS rendering of the story, highlighting LLM stamina for hyper-custom worlds while noting weaknesses in multimodal auditing and gameplay.
Discusses Andrej Karpathy's ability to explain complex technical concepts clearly, and extends to the importance of Developer Relations (DevRel), emphasizing that this skill becomes even more critical in the AI era. Also shares the experience and insights of a DevRel practitioner.
A single CLAUDE.md file that provides structured behavioral guidelines for Claude Code, derived from Andrej Karpathy's observations on common LLM coding pitfalls, has gained 192k GitHub stars. It aims to prevent over-engineering, reduce wrong assumptions, and enforce simplicity in AI-generated code.
Two Hong Kong students achieved a 5x speedup by adding another loop outside the original automated research framework, without needing a better model or more compute. It is considered one of the most useful papers for ordinary Agent developers.
The article provides a detailed analysis of the llm-wiki method proposed by Karpathy: using LLM to compile personal raw materials into interlinked wiki pages, where users only need to read and ask questions without manual editing. The author compares the limitations of RAG and provides a 30-minute setup tutorial, emphasizing the key role of schema files in system discipline.
The author experiments with self-improving AI loops using Claude and tools like AutoResearch, demonstrating that recursive self-improvement is accessible beyond frontier labs and can automate newsletter busywork.
Introducing the Karpathy Loop concept, where AI automatically executes experimental loops without real-time user interaction, boosting work efficiency by 10x.
Explains Karpathy's reader3 method that uses an LLM as a private tutor to help readers actively engage with books chapter by chapter, turning passive reading into lasting understanding through summarization, Q&A, and note-taking.
At the Sequoia AI Conference, Karpathy spent nearly 30 minutes in-depth sharing the dramatic changes in programming paradigm, including concepts like Software 1.0/2.0/3.0 framework, vibe coding, jagged intelligence, etc., and provided practical advice for founders.