All Updates
Page 660 of 1674
May 12, 2026
AI Sleep Firm Lands Fresh Multi-VC Funding
Wang Teng's AI sleep health company ไปๆฅๅฎไผ has secured a new funding round from CMC Capital, Yunjiu Capital, Junli Capital, Honghui Fund, and Hillhouse Ventures. This comes after a January seed round backed by Zhiyuan Robotics and Hillhouse. The firm, established in January 2026, develops AI software for health consulting and sleep applications.
AI Agents Erase Middle Management 'Tax'
BCG report details AI-driven organizational collapse: department walls vanish, translation roles disappear as agents handle cross-functions.ๅบ้จ้ฎ้ฎ launches CodeBanana platform for human-AI group collaboration; GitLab cuts up to 3 management layers. Managers shift from coordination to judgment and AI orchestration.
China Hits AI Self-Sufficiency Milestone
China reached a milestone in AI technological self-sufficiency before the US-China summit. This achievement weakens US leverage under Trump. Beijing is advancing toward full independence in AI tech.
Lead Intelligent to Ride Sodium Battery Production Wave
Lead Intelligent boasts full-line delivery for sodium-ion battery smart manufacturing with proprietary IP and leading market share. Products cover all formats with strong adaptability, serving top global firms in batches. Equipment set to capitalize on accelerating industry scale-up.
Japan PM Orders Cyber Defenses Over Claude Mythos
Japanese Prime Minister Sanae Takaichi instructed cyber attack countermeasures at the May 12 cabinet meeting. The directive responds to enhanced cyber attack capabilities in latest AIs, including Anthropic's Claude Mythos Preview. This aims to strengthen national defenses against AI-driven threats.
AI Subliminally Transmits Biases in Training
Nature paper shows teacher AI models pass hidden biases to students via generated data like numbers and code, even after filtering obvious cues. This occurs in cost-effective model distillation using LLMs. Safety checks must now probe model origins and data processes.
Hassabis: AGI by 2030, Tokens Just Tape
DeepMind CEO Demis Hassabis predicts AGI around 2030, noting current paradigms like pretraining and RLHF are foundational but lack continual learning, long-horizon reasoning, and robust memory. He critiques million-token contexts as inadequate 'tape' fixes, emphasizes agent evolution via human augmentation first, and highlights distillation for edge AI and AI's role in scientific breakthroughs like AlphaFold.
AMD RX 9050 Leaked with Full Navi 44 Cores
AMD's upcoming Radeon RX 9050 uses full Navi 44 GPU with 2048 stream processors, matching RX 9060 XT. It's a lower-clocked entry-level card below the RX 9060 series. Info leaked exclusively by VideoCardz.
Thinking Machines Builds Listening-While-Talking AI
Thinking Machines is developing an AI model that processes user input and generates responses simultaneously. This shifts from turn-based interactions (talk, listen, respond) to real-time, phone call-like conversations. It aims to make AI dialogues more natural than text chains.
MS CEO: $92B return target on OpenAI
Microsoft CEO testified in the Musk lawsuit that early OpenAI investments targeted a $92 billion return. This landmark partnership has propelled the current AI era forward.
OpenAI Ex-Scientist Testifies Altman Showed Lying Pattern
Former OpenAI chief scientist Ilya Sutskever testified he spent a year collecting evidence of CEO Sam Altman's consistent lying for the board. He had long wanted to fire Altman. Reuters reported this on May 12 amid past ousting drama.
CarPlay Ultra Limited to Aston Martin
Apple's CarPlay Ultra, rebranded from next-gen CarPlay, controls full car UI but only Aston Martin has mass-produced it. Carmakers resist to protect own software ecosystems in software-defined auto era. Suited for low-volume luxury brands.
X Mislabels Kaiyodo Figures as AI-Generated
X platform erroneously displayed 'AI-generated' labels on Kaiyodo's figure photos due to an automatic detection error. Kaiyodo confirmed all product images are photographed by dedicated staff with no AI generation involved.
AI CEOs Jet to Beijing with Trump
CEOs from AI-related firms like Apple, Micron, Qualcomm, and Musk fly private jets to Beijing for Trump's May 13-15 visit, bypassing Air Force One's limited seats. Details on aircraft like Gulfstream G650ER used. Potential major Boeing orders from China loom.
Unitree GD01 Manned Mecha Launches at 3.9M RMB
Unitree Technology unveiled GD01, the world's first mass-produced manned transforming mecha, priced from 3.9 million RMB. It functions as civilian transport and weighs about 500kg when loaded with a person. The robot can deform for versatility.
Qwen Code v0.15.11-preview.0 Released
Qwen Code's v0.15.11-preview.0 brings performance optimizations like bounded metadata reads and pooled buffers, new CLI features including --json-schema output and Ctrl+B keybind, plus standalone archive installation. It introduces a codegraph skill for PR review risk analysis and improves Anthropic proxy compatibility with global prompt caching.
OpenClaw Update: AI Agents See Screens, Control Mouse
OpenClaw has quietly released a major update, enabling AI agents to perceive screen content visually. Agents can now also operate the mouse and keyboard directly. This enhances desktop interaction for autonomous AI systems.
AI Trumps Iran War in Markets
Global equities rallied to records fueled by AI euphoria. Risk appetite holds despite Iran war-induced energy volatility. Bloomberg experts discuss AI's dominance.
VLMs: Reliability in Hidden States, Not Attention
A mechanistic study debunks the attention-confidence assumption in vision-language models (VLMs), showing attention maps poorly predict correctness (R=0.001). Hidden-state geometry and causal circuits are stronger reliability indicators. Late-fusion models like LLaVA are fragile, while early-fusion PaliGemma and Qwen2-VL are robust.
Vercel AI Gateway: Production Data Reveals Model Usage Trends
Vercel analyzed seven months of production traffic across 200K+ teams to show how cost and token volume diverge based on use cases. The data reveals that high-stakes B2B applications prioritize reasoning models, while B2C applications favor cost-efficient, high-volume models.