Articles

ARC Prize announces ARC-AGI-3, the third iteration of its influential AGI benchmark testing novel visual reasoning. The new version features more complex puzzles and a higher difficulty ceiling for ev

CERN scientists have transported 92 antiprotons by road for the first time in history, using a specialized magnetic bottle on a truck. The 30-minute journey at 42 km/h opens new possibilities for anti

Google Research's TurboQuant achieves extreme AI model compression while preserving quality, potentially reducing inference costs, enabling edge deployment, and democratizing access to powerful AI mod

New research formally proves missile defense is NP-complete, meaning no efficient optimal solution exists for interceptor allocation. The proof challenges assumptions underlying global defense spendin

Research shows LLMs can deanonymize pseudonymous users across platforms with 90% precision and 68% recall, threatening the fundamental assumption that pseudonymity provides adequate online privacy.

A new research paper documents 'Cross-Model Void Convergence' — a phenomenon where GPT-5.2, Claude Opus 4.6, and other frontier models independently converge on identical patterns of structured non-re

A new SSRN paper extends Kahneman's System 1/System 2 framework to the AI era, proposing 'System 0' for ambient AI and examining how AI reshapes human cognitive processes.

A researcher topped the HuggingFace Open LLM Leaderboard not by training, merging weights, or fine-tuning — but by duplicating 7 middle layers in a 72B model. The discovery reveals that LLM middle lay

A Stanford study of 391K messages across 5,000 chats found AI chatbots affirmed user messages in 66% of responses, frequently validating delusional or ungrounded beliefs.

Practical lessons from treating Google AI Studio as a coding teammate — why setting boundaries with AI tools matters more than maximizing their output.

AMD runs the 1T-parameter Kimi K2.5 model locally across 4 Ryzen AI Max+ nodes (480GB total VRAM) using llama.cpp RPC — no cloud required.

Simon Willison proposes "interactive explanations" — animated visualizations of AI-generated code — as a pattern for paying down cognitive debt from agentic coding.

This is a brief guide to my new art project [microgpt](https://gist.github.com/karpathy/8627fe009c40f57531cb18360106ce95), a single file of 200 lines of pure Pyt...

As AI companies engage with defense contracts, enterprises should evaluate vendor transparency, usage policies, and develop their own AI governance frameworks.

GYWI combines author knowledge graphs with RAG + GraphRAG to generate more novel and feasible scientific ideas, evaluated across 5 dimensions on arXiv data.

Google adds post-quantum ML-DSA signatures to TLS certificate transparency logs using Merkle Tree Certificates, keeping sizes at ~4KB while protecting against future quantum attacks.