Andrej Karpathy
AI research & education
Former Tesla AI Director, OpenAI founding member. Builds educational AI content and nanoGPT.
Recent activity
-
AI models are transitioning from being chatbots or autocompleters to becoming a programmable layer for digital work, enabling humans to delegate larger "macro actions" and become orchestrators of agents rather than just code writers. This shift is characterized by the increasing use of context windows and multimodal models, which can automate complex information transformations and processes previously impossible with classical programming. As a result, the profession is being refactored to focus on tasks that can be verified and rewarded, such as coding, testing, and benchmarking. AI summary
Read more → -
The 2025 LLM year in review saw significant paradigm changes, including the emergence of Reinforcement Learning from Verifiable Rewards (RLVR), which enabled LLMs to develop "reasoning" strategies through training against verifiable rewards across various environments. This shift led to the development of longer, more capable LLMs, and raised questions about the nature of LLM intelligence, which is distinct from human intelligence and cannot be understood through an animal lens. The year also saw the rise of LLM apps, such as Cursor, which bundle and orchestrate LLM calls for specific verticals, and the emergence of LLM agents, like Claude Code, which demonstrate extended problem-solving capabilities. AI summary
Read more → -
To minimize exposure to toxic chemicals in drinking water, consider installing a Reverse Osmosis point-of-use filtration system with a remineralization post filter, and consider whole-home water filtration for added protection. For air quality, install HVAC filters, standalone air purifiers, and avoid combustion sources, combustion byproducts, and synthetic fragrances. AI summary
Read more → -
The author developed an AI model to grade decade-old Hacker News discussions with hindsight, analyzing comments on the front page of December 2015 and comparing them to actual outcomes. The model achieved impressive results, with some comments being highly prescient and others completely wrong. The author plans to release the code and data on GitHub, allowing others to reproduce and analyze the results. AI summary
Read more → -
The optimization pressure driving animal intelligence is fundamentally distinct from that of artificial intelligence, with animal intelligence optimized for survival, self-preservation, and social dynamics, whereas LLMs are optimized for problem-solving and seeking user validation. This difference in optimization pressure leads to vastly different computational substrates, learning algorithms, and present-day implementations. As a result, LLMs are more akin to a "first contact" with non-animal intelligence, with humans still grappling with understanding and predicting their behavior. AI summary
Read more → -
The concept of verifiability in AI programming is key to automation, as it enables direct optimization and reinforcement learning, allowing neural networks to excel in tasks that can be reset, efficiently attempted, and rewarded. Verifiable tasks are more amenable to automation in the new programming paradigm, while non-verifiable tasks rely on weaker means like imitation or generalization. This distinction is driving progress in Large Language Models (LLMs). AI summary
Read more → -
Researchers in the field of Large Language Models (LLMs) are being challenged by Sutton, who argues that LLMs are not "bitter lesson pilled" and are instead heavily reliant on human-generated data, which can lead to bias and a lack of generalizability. Sutton proposes an alternative architecture that focuses on reinforcement learning and intrinsic motivation, rather than relying on human-designed reward functions and pretraining. This approach is inspired by the way animals learn and interact with their environment. AI summary
Read more → -
Here's a summary of the article in 3 plain sentences for a developer/AI-ML audience: The author, a non-web developer, built a vibe-coding project called MenuGen, which generates images of menu items from a restaurant's image. The project leveraged OpenAI and Replicate APIs, but faced challenges with rate limiting and outdated documentation, requiring the author to debug issues manually; these issues were eventually resolved with the help of Cursor+Claude and online resources. The author deployed the app on Vercel, but encountered issues with environment variables not being recognized, which was resolved by manually adding API keys to the project settings. AI summary
Read more → -
Large Language Models (LLMs) have flipped the traditional script on technology diffusion by providing disproportionate benefits to individuals, with 400 million weekly active users utilizing them for various tasks, despite being less impactful in corporate and government realms. This reversal is attributed to LLMs' shallow yet versatile capabilities, which enable individuals to access a broad range of knowledge and expertise without needing specialized technical expertise. In contrast, organizations face more complex problems, coordination challenges, and constraints, making it harder for them to adopt and utilize LLMs effectively. AI summary
Read more → -
A developer/AI-ML audience can analyze the findings of this sleep tracking experiment by noting that the Oura and Whoop sleep trackers produced high-quality, correlated sleep scores, while the 8Sleep tracker showed a more limited range of scores. The correlation analysis revealed that Oura and Whoop had the strongest correlation with each other (0.65), while the Apple Watch-based AutoSleep tracker showed little correlation with other trackers. AI summary
Read more →