$40B+ raise at ~$100B valuation (3x last year's mark), defense tech/AI drones, talks ongoing per Reuters
A major policy letter titled 'Open Weights and American AI Leadership' was released July 24, 2026, co-signed by 25 US companies including NVIDIA (Jensen Huang's first ever X post), Meta, Microsoft, IBM, Mistral AI, Hugging Face, and Mozilla, arguing open-weight AI models expand American AI leadershi
NVIDIA CEO Jensen Huang made his first-ever X post on July 24, 2026, sharing the 'Open Weights and American AI Leadership' letter signed by 25 US companies including NVIDIA, defending open-weight AI models as critical to US AI leadership.
BAAI presenta AREX, una famiglia di agenti deep research con auto-miglioramento ricorsivo. Alterna ricerca interna e un loop esterno di self-improvement che verifica le risposte vincolo per vincolo. Un modello 122B-A10B MoE supera baseline comparabili su BrowseComp, WideSearch e Humanity's Last Exam
Microsoft presenta un framework open-source per addestrare agenti harness-based (Claude Code, Codex, OpenClaw) end-to-end via RL. Usando solo centinaia/migliaia di task, OpenForgeClaw raggiunge 31.7 pass^3 su ClawEval e OpenForgeGUI 72.3 su WebVoyager, superando baseline aperte di dimensioni simili
Propone una sliding window + attention sink per il draft head di MTP/NEXTN, riducendo del ~99% il KV cache del draft a 1M di contesto. Su Qwen GDN-MoE 35B/122B e Mamba2-hybrid 120B, taglia il costo per step del 28-44% senza perdita di qualità, essendo il full-attention target a decidere i token acce
Google presenta JAXBench, benchmark TPU-native con 50 workload JAX. Con Gemini 3 Flash, la documentazione curata alza la correttezza per-sample dal 5.8% al 37.3%, risolvendo 48/50 benchmark con speedup 1.28x. Autocomp raggiunge 1.36x geomean su XLA e 1.60x sui kernel ottimizzati a mano.
Framework training-free che accelera i Diffusion LLM del 53.19x su MBPP per sequenze lunghe, e fino a 105.02x combinato con KV-Cache. Introduce Dynamic Contiguous Verification che neutralizza il Joint Probability Dependence Error (JPDE), consentendo decoding parallelo anche in regime di confidenza m
Algoritmo inference-time che con feedback identifica il token esatto dove avviene l'errore, tronca la traiettoria e rigenera riusando il prefisso valido. Con Qwen3-8B su LiveCodeBench raggiunge pass@64 del 71.0% generando la metà dei token rispetto a independent sampling (360.4k vs 735.0k).
OpenAI launched a physical AI keypad peripheral designed for coders, offering quick access to AI completions and code generation directly from a dedicated hardware input device. Reviewed by TechCrunch as fun for some coders and mystifying to others.
GitHub made its Copilot cloud agent integration for Linear generally available, allowing teams to assign Linear issues to an asynchronous coding agent that produces draft pull requests for review.
OpenAI launched a U.S. rollout letting eligible ChatGPT users connect health records and supported wellness data for more contextual health conversations, powered by GPT-5.6 Sol on paid plans.
Bluesky's AI assistant Attie expanded to let users ask questions about news, trends, and conversations across Bluesky and other AT Protocol apps, acting as an open social research tool.
Pushary launched native iPhone and Android apps that put AI agent approval requests on your lock screen. One command connects Claude Code, Codex, Cursor, Gemini CLI, Hermes, and Claude Cowork to your phone.
Fluree AI gives every app and AI agent the same trusted context from company data, providing cited verifiable answers from a live data layer with permissions checked on every request, connecting to MCP-ready agents.
South Korean President Lee Jae Myung issued the 'San Francisco AI Declaration' at the NVIDIA AI Summit, joined by Jensen Huang, Sam Altman, Hock Tan, and Dario Amodei. Korea outlined a multitrillion-dollar push into AI semiconductors, hyperscale data centers, and physical AI. NVIDIA and KAIST announ
Intel is planning layoffs in its Data Center Group (DCG) as part of CEO Lip-Bu Tan's broader effort to create a 'more focused and efficient' company. The layoffs target the division focused on server CPUs, AI chips, and data center architecture. Intel declined to disclose the number of affected empl
Deep-dive analysis of AMD's Helios rack-scale strategy following Advancing AI 2026, examining the financial opportunity AMD is chasing with its integrated rack-scale approach, the competitive dynamics against NVIDIA Vera Rubin NVL72, and AMD's roadmap through Helios 500 and Helios 600 systems extend
AI4I, the Italian Institute for Artificial Intelligence for Industry, saw its SUK (System for User Knowledge) technology transfer platform featured in a full-page article on Il Sole 24 Ore. The platform connects 140 qualified AI providers, 59 innovative startups, and 70 SMEs through a structured mat
Gianluigi Greco, president of Italy's National Committee for AI, told AI4Business that public debate on AI is polarized between unjustified alarmism and unrealistic expectations, urging strategic decisions anchored in scientific rigor rather than hype.
The Food Valley industrial district in Emilia-Romagna is adopting artificial intelligence not only to innovate food production but increasingly to conquer new international markets, positioning the region as a global laboratory for AI-driven agrifood competitiveness.
$100M at $1B valuation, computer-use AI agents, seed stage, co-founded by Reid Hoffman and Mark Pincus, $50M in signed contracts, sector: AI agents/enterprise automation
$52.5M via WLD token sale with 12-month lockup, lead investor: Pantera Capital, joined by Bain Capital Crypto, Susquehanna Crypto, Selini Capital, sector: biometric AI/identity verification