22 Sep 2025

AI News Digest

🤖 AI-curated 8 stories

Today's Summary

Nvidia and Abu Dhabi’s TII are teaming up to open the Middle East’s first Nvidia AI Technology Center, pushing the UAE closer to its goal of becoming a global AI powerhouse. Meanwhile, Google is making waves by folding its Gemini AI into Chrome, which means we could soon be outsourcing our web browsing chores to AI. And in the AI hardware world, Cerebras is on the brink of a massive funding round, highlighting the continued excitement around innovative players challenging the GPU giants.

Stories

Nvidia and Abu Dhabi’s TII open joint AI & robotics research lab — first Nvidia AI Technology Center in Middle East

Abu Dhabi’s Technology Innovation Institute (TII) and Nvidia launched a joint AI and robotics research hub on Sept. 22, 2025 — the region’s first Nvidia AI Technology Center. TII will get access to Nvidia’s Thor edge GPUs and will host collaborative teams to develop next‑generation AI models and robotics platforms (humanoids, quadrupeds, robotic arms) with target use cases in climate, energy, genomics, transportation and logistics. The partnership accelerates the UAE’s push to become a global AI player, gives TII more compute for local model training, and signals continued geopolitical and commercial momentum for major chipmakers expanding research partnerships outside the U.S. and Europe.

Mycroft emerges from stealth with agentic AI security platform and $3.5M seed

Mycroft Technologies launched from stealth on Sept. 22, 2025, unveiling an agentic AI‑native security and compliance platform that promises continuous, autonomous monitoring, remediation and audit readiness across cloud infrastructure and endpoints. The Toronto‑based startup announced a $3.5M seed round led by Luge Capital (with participation from Brightspark, Graphite Ventures and others). Mycroft positions its product as a virtual CISO/GRC/IT‑ops team for startups and scaleups — a timely offering as organizations seek automated security workflows powered by agentic AI.

DeepSeek’s R1 model published in Nature — paper reveals $294K training bill and peer‑reviewed details

DeepSeek published a peer‑reviewed article in Nature describing R1, an open‑weight reasoning‑focused LLM. The paper discloses technical details and — for the first time — a training cost estimate of roughly $294,000 (plus earlier base‑model expenses) and explains R1’s reinforcement‑learning approach for improved step‑by‑step reasoning. The Nature publication marks an uncommon instance of a major production LLM undergoing independent peer review, which could raise expectations for greater transparency and reproducibility in high‑impact model releases and renew debate about resource efficiency, export‑controlled hardware, and model provenance.

Generative Data Refinement (GDR): Google/DeepMind paper proposes using generative models to 'clean' unusable training data

Researchers (affiliated with Google DeepMind) posted an arXiv preprint introducing Generative Data Refinement (GDR), a framework that uses pretrained generative models to rewrite or sanitize mixed‑quality data (removing PII, toxicity or other problematic content) while preserving useful signal. The authors report strong empirical gains versus detector/rule‑based pipelines and argue GDR can recover large volumes of otherwise discarded tokens — a potential solution to an emerging 'data drought' for frontier model training. The idea is timely because labs are increasingly concerned about the long‑term supply of high‑quality training data and about privacy/safety tradeoffs when using more aggressive data sources.

Read more → Business Insider (reports on arXiv preprint: arXiv:2509.08653)

Oracle names two insiders as co‑CEOs in surprise leadership shakeup

Oracle on Sept. 22, 2025 appointed Clay Magouyrk (head of Oracle Cloud Infrastructure) and Mike Sicilia (head of industry cloud apps and AI solutions) as co‑CEOs, with longtime CEO Safra Catz moving to executive vice chair. The move preserves executive continuity while elevating leaders directly tied to Oracle’s cloud and AI business lines — a signal that the company is leaning into AI-driven infrastructure and vertical applications as strategic growth engines. Markets reacted modestly (shares dipped ~1% premarket), but the change reduces uncertainty about Oracle’s AI roadmap and could accelerate product and sales motions tied to large AI deployments.

Axios: AI chipmaker Cerebras is nearing a $1B private round as IPO talks continue

Axios (Pro) reports Sept. 22, 2025 that AI‑chip maker Cerebras Systems — which has previously been linked to IPO plans — is nearing a roughly $1 billion private funding round (reported to be led by Fidelity) at a valuation reportedly north of $8 billion. If completed, the deal would bolster Cerebras’s war chest for scaling AI compute hardware and could reshape timelines and terms for any public offering; it also underscores continued investor appetite for companies that challenge incumbent GPU suppliers in the high‑performance AI compute market.

Google folds Gemini into Chrome, bringing AI features (and agentic tasks) to mainstream browsing

Google has integrated its Gemini generative-AI into Chrome (desktop and parts in mobile), adding a Gemini chatbot button, an upcoming ‘AI Mode’ in the Omnibox, page-aware Overviews and planned agentic features that can automate multi-step web tasks (e.g., shopping or booking). Why it matters: embedding Gemini into Chrome pushes AI-assisted research, summarization and basic automation to millions of everyday users — accelerating adoption of AI browser tools and changing how people search and interact with web content. Impact: broader access to in‑browser AI will raise productivity for many users, increase competition among AI-focused browsers, and surface new privacy/usability tradeoffs as agentic capabilities roll out.

Perplexity opens its AI browser Comet to Pro users in India, expanding access to automation features

Perplexity is making its AI browser Comet available to Pro (and Max) subscribers in India starting Sept. 22, 2025, and is granting a large batch of waitlisted users access. The Comet release includes the Comet assistant for automating tasks such as booking meetings and sending emails. Why it matters: regional expansions like this broaden real-world user testing and adoption for AI-first browsing experiences, giving users browser-native AI assistants and intensifying competition with Google and other AI browser entrants.