Annonces

MiniMax's VTP framework improves image generation by replacing pixel-reconstruction tokenizer training with semantic understanding, achieving a 65.8% FID improv

Mistral AI rebranded Le Chat as Vibe, unifying its AI agent to handle both enterprise workflows and software development in a single platform.

PraisonAI claims 14-microsecond agent instantiation and support for 24 LLMs, aiming to unseat LangChain and CrewAI, but lacks production proof.

A multidisciplinary team proposed a testable framework using integrated information theory to determine if AI can become sentient, with implications for ethics

AgentScope's PawBench benchmark, testing 9 models on 150 tasks, reveals that harness design can swing scores by over 11 points, showing the framework matters mo

Researchers at UC San Diego and Aether AI Lab developed StructAgent, a state-centered framework that improved AI agent success on desktop tasks from 27% to 47%

Sakana AI's brain-like neural network study shows the key component flips between MNIST and CIFAR, warning against single-benchmark evaluation.

Google Vids adds Gemini Omni for text-to-video and personal avatars, letting users star in videos without a camera, with SynthID watermarks for authenticity.

Parallel agents promise speed and specialization but fail when architecture lacks task decomposition and state isolation, per Kimi Agent Swarm analysis.

Google released three new Gemini models on July 21, a cheaper Flash, a faster Lite, and a government-only cyber variant, to improve token efficiency and laten

Poolside's 33B-parameter Laguna XS 2.1 beats 137B models on coding benchmarks, achieving 63.1% on SWE-bench Multilingual without hardware changes.

Sakana AI released Fugu-Cyber, a multi-agent cyber model matching frontier benchmarks, but warns it alone cannot fix enterprise security.

OpenCode slashes AI coding costs by 80% with new Gemini 3.5 Flash Lite model, alongside Gemini 3.6 Flash, aiming to make AI coding default for teams.

A practical guide for professionals explains how to structure prompts for Claude to avoid halving answer relevance.

Leaked testing of Gemini 3.6 flash with no sign of expected 3.5 pro hints at a roadmap reshuffle at Google DeepMind.

UPI Help's new hierarchical agent architecture cuts per-call context tokens by 96% while maintaining perfect routing accuracy, using a tree of skills and LIFO s

xAI's Grok now runs scheduled and email-triggered tasks without supervision, acting as a persistent personal assistant on grok.com and mobile apps, with email t

Google prioritizes reach over accuracy in health AI, backing Med-Gemini's 91.1% exam score with millions of free diabetic retinopathy screenings in India, Thail

EU gave Google until July 2027 to open Android to rival AI assistants under DMA, but denied Apple same grace for Siri, highlighting Google's superior regulatory