📰 News
AI News — page 18
Funding, launches, analysis, and interviews.
Anthropic ends Conway agent test July 24 as public preview looms
Anthropic will discontinue its internal Conway agent experiment on July 24, signaling either a wind-down or preparation for broader public access to the always-on Claude container.
NVIDIA Unveils Cosmos 3 Edge and AI Agent Tools at SIGGRAPH 2026
NVIDIA launched Cosmos 3 Edge, a 4-billion-parameter world model for real-time on-device physical AI, alongside Model Context Protocol integrations for creative applications at SIGGRAPH 2026.
China's Open AI Strategy Gains Ground as US Models Stay Proprietary
Chinese AI companies are releasing competitive open-weight models while US firms maintain closed systems, potentially shifting global AI adoption patterns toward Chinese alternatives.
Anthropic tests Penlight for live clinical transcripts and AI research
Anthropic is developing Penlight, a Claude tool that records patient visits, provides real-time transcription, and connects conversations to medical literature during clinical appointments.
Augment Code challenges lean AI coding harness approach with semantic retrieval
Augment Code's VP of Engineering argues semantic context engines outperform grep-based approaches like Claude Code for private codebases, claiming 33% better token efficiency.
Anthropic tests new Projects placement in Claude desktop redesign
Anthropic is testing a revamped Claude desktop layout that moves the mode switcher to the top and integrates Projects directly into the sidebar navigation.
TSMC commits $100 billion more to Arizona fabs as AI demand surges
Taiwan Semiconductor Manufacturing Co. is adding $100 billion to its Arizona investment, bringing total commitment to $265 billion as CFO cites multi-year AI megatrend driving unprecedented chip demand.
LoRA Speedrun Creates Public Leaderboard for Fine-Tuning Speed Records
Open-source project launches competitive leaderboard tracking wall-clock fine-tuning times on frozen hardware, with current record at 22 seconds for Qwen2.5-1.5B on GSM8K.
Alibaba's Qwen3.8 Goes Open-Weight With 2.4 Trillion Parameters
Alibaba announced Qwen3.8, a massive 2.4-trillion-parameter model set for open-weight release, while Moonshot AI's Kimi launches CLI tools and plans Hong Kong IPO.
S1-Omni Unifies Scientific AI Into Single Multimodal Reasoning Model
Researchers released S1-Omni, a unified multimodal AI model that handles scientific understanding, prediction, and generation across domains, outperforming GPT-5.5 and Gemini-3.1-Pro on most benchmarks.
Loopie Models Challenge Transformer Scaling Laws with Novel Loop Architecture
Researchers introduce Loopie, a series of looped Transformer models that outperform traditional scaling approaches by repeatedly processing information through the same network layers.
Frontier AI models score highly on business case analysis benchmark
New research shows leading AI models already perform well on complex business reasoning tasks that mirror white-collar analytical work, with capabilities improving substantially over two years.
Samsung Health threatens data deletion unless users consent to AI training
Samsung Health app now requires users to consent to AI training using their personal health data or face permanent deletion of their stored information.
Apple sues OpenAI over alleged trade secret theft by former engineer
Apple filed a lawsuit against OpenAI claiming former engineer Chang Liu used an authentication bug to steal confidential hardware files after joining the AI company in January 2026.
Apple's trade secrets lawsuit reveals OpenAI's alleged culture of misconduct
Apple's 41-page complaint against OpenAI details extraordinary allegations including employees joking about unauthorized system access and claims the AI company's culture normalizes trade secret theft from the top down.
Anthropic discovers hidden 'J-space' where Claude puzzles through problems internally
Anthropic researchers found a hidden computational space inside Claude where the AI model processes internal thoughts and commentary before generating responses, offering new insights into how large language models reason.