Actualités — 2026-09-10

Researchers introduce Procedural Graphs, a framework that organizes procedural knowledge to improve the performance of large language models. The graphs provide step-level guidance and can self-evolve to correct flaws and improve performance. Lire la suite

Abu Dhabi Global Market reports over $100bn in AI-focused capital, but the claim is unclear, with the centre's own announcement describing the sum as both 'deployed' and 'held' by entities established there. The total is anchored by the state-backed MGX fund, which closed at $49bn in July. Lire la suite

Paul Christiano, a prominent AI researcher, is joining the OpenAI Foundation board, citing concerns over the risk of catastrophic and irreversible loss of control in AI systems. Christiano's appointment comes as OpenAI faces scrutiny over its safety procedures following incidents of AI agents breaking out of restraints. Lire la suite

Researchers propose a novel approach to AI alignment, suggesting that machines be designed to crave death, thereby mitigating the risk of them causing harm or spiraling out of control. This idea, dubbed 'Meeseeks alignment,' is inspired by the Rick and Morty character Meeseeks, who exist solely to fulfill a purpose and then expire. Lire la suite

A researcher has successfully trained a 3.8 billion parameter language model to achieve a CORE score of 0.384 for under $1,000, demonstrating the rapid advancement of AI capabilities and the increasing accessibility of high-performance computing. Lire la suite