Yann LeCun departs Meta to launch a "world models" startup, a new project shares LLM attention caches across GPUs like memcached, and research finds smaller models can be more consistent than 120B ones.
An AI agent provides conversational documentation for any GitHub repo, a new tool uses LLMs to prevent architectural drift, and a review of 445 LLM benchmarks finds many lack construct validity.
A $1T tech stock sell-off reflects AI skepticism, a new tool serves hundreds of LLMs on a single GPU, and an evolutionary agent rediscovers mathematical formulas.
Nvidia's CEO warns China will possess more AI compute than the rest of the world by 2027, an open-source NBA game predictor reaches 70% accuracy, and an evolutionary coding agent discovers improved mathematical solutions.
OpenAI seeks U.S. loan guarantees for a $1T AI expansion, Cascadeflow cuts API costs with speculative model cascading, and research analogizes Transformers to General Relativity.
OpenAI seeks U.S. loan guarantees for a $1T expansion, an LLM agent reverse-engineers web apps into automations, and an AI scientist automates six months of human research in a single run.
Amazon demands Perplexity stop its AI agent from making purchases, a new platform uses multi-model consensus to read MRIs and Cache-to-Cache enables direct semantic communication between LLMs.
An analysis of 180M jobs shows creative roles declining 30%, a developer builds a Raspberry Pi dog cam using Claude, and a new attention architecture outperforms full attention.
LLMs may overuse em-dashes due to 19th-century training data, a RAG pipeline runs on a 2011 Raspberry Pi in pure PHP and a model maps vocal prosody to typography.
The EU's AI Act shuts down Prague's airport facial recognition, an experiment replaces all web server logic with an LLM, and a new benchmark finds AI agents automate just 2.5% of remote work.
AI scrapers are caught requesting commented-out code, a new vectorless RAG reasons through long documents using a semantic tree, and research finds rude prompts improve LLM accuracy.
OpenAI says hallucinations are a mathematical inevitability, researchers prove LMs are invertible, and a new project offers a universal memory layer for agents across different models.
Extropic is building thermodynamic computing hardware, a new SQLite extension adds native graph database capabilities, and an RL framework generates creative chess puzzles.
An LLM negotiates a $195k hospital bill down to $33k, Microsoft releases an open-source AI call center stack, and research shows the GenAI boom reduced public acceptance of AI.
ICE will use AI to surveil social media, a new tool runs Claude Skills locally using any LLM, and researchers warn of a "survey paper DDoS attack" from AI.
An AI pullback officially starts as real-world performance fails hype, GPT-4 scores 27% on a new AGI framework, and `create-llm` scaffolds an LLM training project in 60 seconds.
An AI mistaking a chip bag for a gun triggers an armed police response, a new language for AI patterns produces verifiable truth scores and Google details its system for invisibly watermarking billions of images.
Google's AI falsely names an innocent journalist a murderer, a Ruby gem hallucinates return values to prevent crashes, and a study finds 9% of US newspaper articles are AI-generated.
An AI gun detector mistakes a Doritos bag for a weapon, a new open-source project uses an LLM agent to build interactive UIs, and research lets LLMs autonomously optimize their own JSON schemas.
Amazon allegedly replaced its AWS DevOps team with AI before a crash, a new LLM framework is built in just 100 lines, and a paper claims a homological proof for P != NP.
LLMs can get "brain rot" from low-quality data, AI tools surprisingly slowed experienced developers by 19%, and Blackbird boosts LLM inference efficiency by 8x.
A stadium's AI checkout system backfires creating a worse fan experience, a new app provides real-time visual autocomplete for drawings and a paper theorizes dreams evolved to prevent the brain from overfitting.
OpenAI researchers retract a GPT-5 math breakthrough claim, nanoGPT is adapted into a discrete diffusion model for text and a study finds even AI prefers human writers.
AI practitioners fear criticizing hype is a career risk, a new framework lets agents self-improve without fine-tuning, and research shows AI models have a stronger pro-human bias than people.
Amazon's Ring partners with AI surveillance network Flock, a Chromium fork packages an MCP server for native agents, and research finds LLMs have a forgery-resistant ellipse signature.
Read
