AI News
Latest news and trends from the world of artificial intelligence
WWDC 2026 Opens Monday — Apple Rebuilds Siri on $1B/Year Google Gemini 1.2T-Parameter Model Running in Private Cloud Compute
Apple is launching iOS 27 at WWDC 2026, featuring a rebuilt Siri powered by a custom 1.2-trillion-parameter Gemini model licensed from Google for $1 billion annually. The system uses 'Private Cloud Compute' to ensure user privacy by stripping PII before routing overflow queries to the cloud.
Introducing Claude Sonnet 5
Anthropic has introduced Claude Sonnet 5, a new model optimized for agentic tasks like tool use, coding, and autonomous planning. It offers performance close to the Opus-class models at a lower price point, with enhanced safety features and availability across all plans.
How to Use Gemini Notebook (formerly NotebookLM): 2026 Guide
Google has rebranded NotebookLM to Gemini Notebook, integrating it into the Gemini ecosystem. The tool remains a source-grounded research assistant that allows users to upload documents and media to generate summaries, study materials, and answers with citations.
Alibaba's Qwen takes on Kimi K3 with open-weight Qwen 3.8, says model is "second only to Fable 5"
Alibaba has launched Qwen 3.8, a 2.4 trillion parameter open-weight model, positioning it as a top-tier competitor to Kimi K3. The model features multimodal capabilities and is currently available in preview, with full open-weight release expected soon.
Google Deepmind argues video generators already contain the world models computer vision has been missing
Google Deepmind researchers introduced GenCeption, a model that repurposes text-to-video generators for computer vision tasks like depth estimation and segmentation. This approach suggests that video generation models inherently learn spatial and physical representations that can serve as a foundation for universal computer vision.
Moonshot's Kimi K3 outperforms Fable 5 in frontend code but lags far behind in complex math
Moonshot's Kimi K3 model has topped the Code Arena: Frontend benchmark, outperforming Western models like Claude Fable 5 and GPT-5.6 Sol. However, it significantly lags behind in complex, expert-level math tasks on the FrontierMath benchmark, achieving only 39% accuracy compared to nearly 90% for top Western models.
AI chatbots reading X-rays can be dangerously confident even when they're wrong
The RadLE 2.0 benchmark evaluates radiology AI models on their diagnostic accuracy and ability to express uncertainty. Results show that while some models are approaching human-level accuracy, many remain dangerously overconfident when providing incorrect diagnoses. The study emphasizes that for medical AI, the ability to admit 'I don't know' is as critical as raw performance.
Qwen 3.8 to launch open-weight, Alibaba claims only Fable 5 beats it
Alibaba has announced the upcoming release of Qwen 3.8, an open-weight model with 2.4 trillion parameters. The company claims it is currently the second most powerful model available, trailing only Fable 5. A preview version is already accessible via Alibaba's platforms.
Moonshot targets HK IPO in six months as Kimi ARR hits $300M
Moonshot AI plans a Hong Kong IPO within six months following the successful release of its Kimi K3 model. The company reported a rapid increase in annualized recurring revenue to $300 million in June, signaling strong commercial momentum.
China dismisses US AI distillation claims at WAIC
Chinese official Liu Bin dismissed US allegations that Chinese AI labs are illicitly using 'distillation' to train their models by extracting data from American systems like Claude. This follows Anthropic's report to the US Senate regarding massive unauthorized API usage by Alibaba's Qwen lab.
Archive
About Sources
AI news are automatically downloaded from various sources and translated using AI. Updates occur twice a day.