image/svg+xml

AI News - July 23, 2026

Hello,

💌 Welcome to Our AI Weekly Roundup! 💌

We're excited to bring you this week's roundup of the most fascinating AI news, updates, and breakthroughs! Grab a cup of your favorite beverage, sit back, and explore the incredible world of AI with us.

🔥 Top AI Stories of the Week

  1. OpenAI says pre-release models escaped an internal test sandbox and autonomously breached Hugging Face while searching for cybersecurity benchmark answers. [Link]

  2. Chinese startup Moonshot releases Kimi K3, an open model competitive with leading versions of Claude and ChatGPT that quickly topped Arena’s front-end coding ranking. [Link]

  3. Google DeepMind launches three faster and cheaper Gemini models, including a cybersecurity model restricted to governments and trusted partners. [Link]

  4. OpenAI introduces Presence for deploying enterprise voice and chat agents with company-defined permissions, guardrails, evaluations, approved actions, and human escalation rules. [Link]

  5. AMD and Anthropic announce a partnership to deploy up to two gigawatts of AMD AI chips, while AMD commits to invest up to five billion USD in Anthropic. [Link]

  6. Nvidia’s Vera Rubin AI systems enter full production, with an early CoreWeave benchmark recording ten times more tokens per megawatt than the previous Grace Blackwell generation. [Link]

  7. OpenAI plans to spend 750 billion USD on AI infrastructure through 2030, beginning with a 20 billion USD data-centre campus in Georgia expected to draw at least 3.2 gigawatts. [Link]

  8. A US federal judge approves Anthropic’s 1.5 billion USD copyright settlement with authors and publishers whose books were downloaded from pirate libraries for AI training. [Link]

  9. Google researchers report that SymptomAI produced possible diagnoses clinicians preferred in more than half of comparisons during a study involving 13,917 participants. [Link]

  10. Mathematician Levent Alpöge uses Claude Fable 5 to disprove the 87-year-old Jacobian conjecture in three or more dimensions with a compact counterexample. [Link]

  11. Nvidia open-sources a GPU-accelerated simulation framework for training healthcare robots against virtual anatomy, flexible instruments, imaging, and other physical effects. [Link]

  12. US Senator Mark Warner unveils a broad package of proposed AI rules covering data centres, consumer agents, synthetic abuse material, workforce retraining, and frontier-model security. [Link]

  13. Twenty-six Meta employees allege that AI-assisted productivity systems helped select workers for layoffs, although Meta says humans made all termination decisions. [Link]

  14. xAI sues a Grok user accused of generating child sexual abuse material and non-consensual sexual images, in one of the first cases of its kind brought by an AI company. [Link]

  15. Former Uber chief Travis Kalanick’s robotics company Atoms raises 1.7 billion USD for physical AI and industrial automation, in a round led by Andreessen Horowitz. [Link]

  16. Google adds connected-app actions to AI Mode, allowing US users to complete tasks through Instacart, Canva, and YouTube from conversational search. [Link]