Annonces

LTM and Cognition launch BlueVerse RightLogic, a managed service using Devin AI to automate vulnerability remediation for finance and insurance, claiming 80% CV

xAI released Grok 4.5, its most capable model, across mobile, desktop, and enterprise platforms, with integrated add-ins for Microsoft and Google apps.

AgentRCA uses a digital twin and LLM to diagnose industrial faults without labeled data, matching supervised baselines and producing transparent reasoning.

Cursor launches a 649 INR/month India plan with UPI and Cursor Router, an intelligent model routing system, available from July 28, 2026.

Nvidia launched Cosmos-H-Dreams, a real-time generative simulator for surgical robotics, distilling the Cosmos-H-Surgical-Simulator into a causal student model

CHARM: A multimodal graph foundation model uses hierarchical context modeling and a modality-aware encoder to achieve zero-shot transfer across graphs with text

OpenForgeRL is an open-source framework that decouples training from harness inference, enabling RL training of any harness-based agent with hundreds to a few t

Cognizant and Anthropic expand partnership: 30,000 staff trained on Claude, contract reviews cut 40%, and underwriters save eight hours weekly.

New benchmark DDB tests agent understanding of screen transitions; best models fail 35% of simple ordering tasks, highlighting verification blind spot.

BAAI open-sources AREX, a family of recursive self-improving deep research agents that alternate evidence gathering with constraint-wise auditing, beating large

SceneActBench tests 11 VLM configurations on 5 multi-object 3D tasks; scores top out at 50.2, with no model consistent across all tasks.

France's Pionniers de l'IA program expanded to 51 AI startups after a second wave of 28 winners, with health and industrial robotics receiving the most funding.

US charges Sam Tunick for using a duress password to wipe his phone at Atlanta airport, invoking a rarely-used federal statute against destroying property to pr

Alibaba released Qwen-Music, a music generation model that plans melodies first (Melody-CoT), beating Suno V5 and MiniMax Music 2.6 on 13 of 16 objective metric

ClinPRISM, a multimodal LLM for irregular clinical time series QA, achieves SOTA with 4B parameters, 16 tokens, and 0.15s latency.

Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model matching US frontier performance, intensifying the US-China AI rivalry and exposing div

NYU researchers launched AskChem, a claim-centered chemistry search index holding 2.4M claims from 147K papers, with DOI-level provenance for scientists and AI

Alibaba Tongyi Lab's Qwen-UI-Agent claims state-of-the-art mobile GUI scores (97.5% AndroidDaily), per its July 30, 2026 report.

FrontisAI open-sources OpenMLE and Frontis-MA1, a 35B meta-evolution agent that hits 71.21% on MLE-Bench Lite from one RTX 4090, beating GPT-5.5 + Codex.

CASIA's PhiZero world model reasons about future states in a self-learned "physical language" before rendering video, instead of predicting pixels directly