x-karpathy

Andrej Karpathy announced he has joined Anthropic, expressing excitement to return to R&D at the frontier of LLMs while noting he remains passionate about education and plans to resume that work in the future.

Andrej Karpathy suggests prompting LLMs to output responses as HTML for richer visual formatting, and argues that while audio is the preferred input, vision (images, animations, video) will become the dominant AI output. He envisions a progression from raw text to interactive neural videos and simulations, highlighting the need for better input methods like pointing and gesturing.

At Sequoia Ascent 2026, Andrej Karpathy argued LLMs enable new horizons beyond speeding up existing tasks, like fully LLM-native apps and knowledge bases impossible with classical code. He also discussed the "jaggedness" of LLM capabilities and the emerging agent-native economy splitting products into sensors, actuators, and logic.

Andrej Karpathy shares a quote he has been citing frequently: "you can outsource your thinking but you cannot outsource your understanding."

Andrej Karpathy praises the film adaptation of Project Hail Mary for staying close to Andy Weir's book in content and tone. He highlights the thoughtful scientific portrayal of alien biochemistry and the central bromance, though notes the movie leans more into superhero pacing and quips than the book's science depth.

Andrej Karpathy notes a common issue with LLM personalization: memory features are often too distracting. A single past query can be treated as a deep, lasting interest, leading the model to mention it excessively and inappropriately over time.

Researchers have developed "talkie," a 13-billion parameter language model trained exclusively on text from before 1931. The project aims to study how language models generalize by creating vintage-style AI that can communicate as if from the past, with questions raised about whether such a model can learn modern tasks like coding.

Researchers have developed a prototype called Flipbook that streams pixels directly from a model to the screen, bypassing traditional web technologies like HTML and layout engines. The system aims to deliver exactly what users want to see without conventional coding requirements.

Andrej Karpathy appeared on the No Priors podcast with Sarah Guo to discuss AI engineering shifts, model capabilities, and potential SETI-like movements in AI research. The conversation covered topics including AI psychosis, AutoResearch, and second-order effects of AI development.

A compromised version of the LiteLLM Python package (version 1.82.8) was briefly available on PyPI, capable of exfiltrating sensitive credentials like SSH keys and cloud secrets. The malicious package affected any project that depended on LiteLLM, though it was only available for about an hour before discovery.

Andrej Karpathy observed that when building his MenuGen app, the hardest part was assembling various DevOps services like payments, authentication, and databases, not the code itself. He looks forward to a future where an AI agent could handle the entire deployment process automatically, eliminating the need for manual service configuration.

A developer spent four hours using an LLM to improve a blog post argument, only to have the same AI convincingly argue the opposite position. This demonstrates LLMs' ability to argue any direction effectively, which can be useful for forming opinions if approached carefully.

A supply chain attack has compromised the popular npm axios HTTP client library with 300 million weekly downloads. Malicious versions install a remote access trojan, though some users may have avoided infection through version pinning or older installations. Security experts warn this is a live compromise affecting one of npm's most depended-on packages.

Andrej Karpathy describes using LLMs to build personal knowledge bases by indexing source documents into a raw directory, then having the LLM compile them into a markdown wiki with summaries, backlinks, and categorization. The system allows for complex Q&A against the wiki and can generate various output formats like markdown files, slideshows, and images, all viewable in Obsidian.

AI empowers people to increase government visibility, legibility, and accountability by processing vast amounts of public data that was previously only accessible to professionals. This includes analyzing spending, legislation, lobbying, and local government decisions. While these tools could potentially be misused, increased transparency may improve democratic societies.

Farzapedia is a personal Wikipedia created by an LLM processing 2,500 entries from diaries, notes, and messages to generate 400 detailed articles. The approach emphasizes explicit, user-controlled data in universal file formats that can be used with any AI system, giving users full ownership and interoperability of their personal information.