中文 ↗RSS ↗

THE NEWSROOM / AI 新闻

The updates.
The context.

Source-linked briefings on models, agents, products and practical AI.

YYSI publication dates appear on the left; original announcement dates appear under each brief.

Search and categories filter this page. Use the page links below for earlier news.

REPORTED

Practical guides AI industry

Shanghai Police: Viral 510,000-Yuan Peace Hotel Bill Was AI-Fabricated Marketing Stunt, 15 Arrested

According to IT Home, Shanghai Huangpu police said a viral 510,000-yuan 'Peace Hotel' bill was AI-generated, per CCTV. A liquor e-commerce company allegedly used AI software to alter the bill's amount and swap in its own products, then spread fake videos via controlled and paid accounts. Fifteen people were detained; four face criminal coercive measures and eleven administrative penalties.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates Microsoft

Microsoft 365 outage exceeds 10 hours, disrupting Japanese government and AI service connectivity

According to IT Home, A global Microsoft 365 outage affecting Outlook, Teams and OneDrive lasted more than 10 hours, with Downdetector logging nearly 130,000 Outlook reports. Microsoft attributed it to servers failing health checks and overloading remaining infrastructure. Japan's GSS government platform was hit, prompting officials to rely on phones and chat. The disruption also affected connectivity components used by OpenAI and Anthropic services.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates AI industry

Alibaba's HappyShrimp AI Music App Launches on iOS and Android

According to IT Home, Alibaba's ATH team released HappyShrimp, an AI music creation app, on iOS and Android after an August 2026 PC beta. It generates full songs end-to-end, including lyrics, composition, arrangement and vocals, plus instrumental tracks. Standard membership costs about 30 yuan monthly for 150 songs; overseas plans start at $14.90. It is the third member of Alibaba's Happy model series.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates AI industry

Axelera AI Unveils Titania Data-Center AIPU Architecture, Production Set for Q3 2027

According to IT Home, Dutch chip startup Axelera AI disclosed the architecture of Titania, its third-generation AI processing unit aimed at data-center inference. The 400W chip pairs two compute chiplets with 128GB of LPDDR6 memory and targets sparse MoE models. Axelera says production is expected in Q3 2027 and that it has booked over $300 million in orders; the company claims large efficiency and cost gains versus current data-center servers.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates OpenAI

Report: OpenAI Raises $15B in New Private Round at $1.25T Pre-Money Valuation

According to IT Home, CNBC reported that OpenAI raised $15 billion in a new private funding round at a $1.25 trillion pre-money valuation, above the roughly $11 billion previously reported. Sources said the round is not yet closed and is expected to finish within weeks, described as opportunistic financing ahead of a potential 2027 IPO. The report also cited Q3 2026 revenue of about $10 billion and an operating loss of $10.9 billion.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Practical guides Anthropic

Flinn AI Sues Anthropic and OpenAI Over Five Neural Network Patents

According to IT Home, Media report: Flinn Artificial Intelligence Technologies LLC filed two patent suits in Delaware federal court on Nov. 12 against Anthropic and OpenAI, alleging infringement of five patents on iterative attention-based neural network training and processing. The Anthropic complaint targets the Claude model family's attention, retrieval, memory, agentic and multimodal features. The claims are allegations; no court ruling has been reported.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Models & research AI industry

New AI-Assisted Algorithm Independently Confirms M87* Black Hole Ring

According to IT Home, A CSIRO-led Australian team published a Nature study describing GenDIReCT, an independent imaging method that uses closure invariants and an AI model trained on non-astronomical images to avoid presupposing a ring shape. Applied to Event Horizon Telescope data from 2017, 2018 and 2021, it consistently recovered a compact ring about 42.2 microarcseconds across, supporting the 2019 M87* image as real astrophysical structure rather than an algorithmic artifact.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates AI industry

SpaceX Plans Over $74B in AI Data Center Campuses in Texas and Arizona

According to IT Home, SpaceX announced plans to invest more than $74 billion in two AI supercomputing data center campuses, according to IT之家 citing the company. One facility near Starbase in Texas would cost over $24 billion and use self-built gas power; a roughly $50 billion Arizona campus in Cochise County still needs county approval and faces a petition with about 2,400 signatures.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates AI industry

FAA Expands AI Air Traffic Advisory Tool to All 22 U.S. Centers

According to Ars Technica, The FAA expanded its $875 million SMART AI advisory system from the Washington, DC region to all 22 air route traffic control centers nationwide, the agency said. FAA claims the tool helped reroute 180 flights and saved about 10 minutes and 80 gallons of fuel per flight in the initial rollout. Controllers may accept or reject recommendations, and the controllers' union says it was not consulted on design or testing.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates AI industry

Meta's Muse AI Agent Opens Phone-Call Feature to Paid U.S. Subscribers

According to TechCrunch, Meta's consumer AI agent Muse now lets paid U.S. subscribers have it place calls to businesses on their behalf, after a limited early-adopter test in September, TechCrunch reported. Meta says only paying users get the feature; plans start at $20 per month. Chief AI officer Alexandr Wang said a Muse email address option will soon reach free users. Rival agents from Google and startup Instinct offer similar calling.

AI summary · Media reportOriginal source ↗Read brief →
SOURCE

Product updates Amazon / AWS

AWS releases SGLang-Neuron for serving models on Trainium chips

Amazon announced that SGLang-Neuron, built by RadixArk with Annapurna Labs, is publicly available through an AWS Deep Learning Container. The out-of-tree plugin lets users serve models and generate reinforcement learning rollouts on Trainium AI chips using the standard SGLang server interface and generation APIs, running on native PyTorch with NKI kernels and supporting prefill-decode disaggregation over EFA.

AI summary · Official sourceOriginal source ↗Read brief →
REPORTED

Agents & tools AI industry

Google Researcher's AI Agent 'Discovery' of Maya Site Was Already Documented

According to The Verge, A Google researcher's personal side project using 48 AI agents to scan NASA lidar data claimed to find an unmapped Maya site in Mexico, but archaeologists noted the site was documented in an academic journal over a year earlier. The incident, reported by The Verge, highlights tensions over amateurs using AI tools without expert verification, and the researcher acknowledged expert knowledge is needed to validate agent findings.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Models & research Google

Google launches Gemini 4 Hydrogen, replacing 3.8 Flash across its products

According to Ars Technica, Google released Gemini 4 Hydrogen, a faster, cheaper model replacing Gemini 3.8 Flash across its products. Google claims it runs up to 2.5x faster than Flash 3.8 in high thinking mode, cuts fabricated claims by over 5x, and more than triples Terminal-Bench code scores. It is available immediately to AI Pro and Ultra subscribers and via API at unchanged Flash pricing, with broader Search integration.

AI summary · Media reportOriginal source ↗Read brief →
SOURCE

Models & research Hugging Face

IBM Releases Experimental Granite 4.1 LLMON Preview Models

IBM researchers released two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, trained to consume and generate LLMON, an LLM-native markup language using reserved special tokens to separate instructions from data. The team reports indirect prompt-injection attack success rates dropping to nearly zero on some benchmarks while general capability is maintained relative to Granite 4.1 Instruct. A GitHub toolkit and JSON converters accompany the release.

AI summary · Official sourceOriginal source ↗Read brief →
SOURCE

Product updates Amazon / AWS

AWS upgrades AgentCore runtime with faster starts and larger sessions

AWS announced updates to the Amazon Bedrock AgentCore runtime, its managed compute layer for AI agents. Each session still runs in an isolated Firecracker MicroVM, but sessions now start up to 15x faster, scale to 4x more compute (up to 8 vCPUs and 32 GB), and run up to 68% more efficiently, AWS says. Sessions start in under two seconds regardless of image size or concurrency.

AI summary · Official sourceOriginal source ↗Read brief →
REPORTED

Practical guides AI industry

Study Finds Gemini, ChatGPT and Perplexity Cite Risky Retail Sites in Shopping Answers

According to The Verge, A study by IP protection firm Podqi, shared with The Verge, tested ChatGPT, Gemini and Perplexity with shopping questions about 151 brands across 6,795 answers. All three cited risky sites—fake retailers, knockoffs, reregistered or dead domains. Perplexity showed risky sites in 15 percent of answers, ChatGPT 8 percent and Gemini 4 percent. Smaller brands were more exposed. The vendors said they are improving safeguards.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates NVIDIA

Jensen Huang Resigns as NVIDIA CEO; CFO Colette Kress Named Interim CEO

According to IT Home, NVIDIA said Jensen Huang has resigned as president, CEO and board member, effective immediately. The board appointed CFO Colette Kress as interim CEO while it searches for a permanent successor among internal and external candidates. Harvey Jones, former Synopsys CEO, becomes the company's first-ever board chair. The memo, obtained by Reuters, credits Huang with pioneering accelerated computing and the modern GPU platform behind AI.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates Google

Google details privacy safeguards for Android XR smart glasses

According to IT Home, Google outlined privacy safeguards for its Android XR smart glasses made with Warby Parker and Gentle Monster, per IT之家. A recording indicator lights whenever footage is saved and disables the camera if it fails checks; Gemini camera frames are deleted immediately and never used for training or ads; faces are blurred on-device before photos reach Gemini; and video clips are capped at 90 seconds.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates Google

Google, Samsung and eyewear partners open preorders for Gemini smartglasses from $395

According to Ars Technica, Google opened preorders for Gemini-powered smartglasses designed with Warby Parker and Gentle Monster and built on Samsung hardware. Warby Parker's three frames start at $395 in the US; Gentle Monster offers six styles from $449 in the US and South Korea. The screenless glasses use voice and a 12MP camera, with a recording light, a physical off switch, no live streaming and no facial recognition.

AI summary · Media reportOriginal source ↗Read brief →
REPORTED

Product updates AI industry

AMD to Take Stake in PsiQuantum, Expand Quantum Infrastructure Partnership

According to IT Home, AMD and quantum-computing firm PsiQuantum said they will expand their partnership to co-design infrastructure for practical quantum computers, with AMD planning to take a stake of undisclosed size. AMD will contribute CPUs, GPUs, adaptive computing and software, while PsiQuantum supplies its silicon-photonics platform. The companies said PsiQuantum drug-discovery work already runs on AMD hardware with more accurate results in covalent inhibitor simulation.

AI summary · Media reportOriginal source ↗Read brief →