AI News

Domestic public fund companies are actively introducing AI intelligent agents, upgrading them from auxiliary tools to digital employees capable of independent execution, applied in areas such as investment research and risk control. The industry is currently in the technology verification phase, with institutions taking a cautious approach; for example, Everbright has established a team to explore in an isolated environment.

Google introduces the TurboQuant algorithm, which uses PolarQuant and QJL technologies to reduce the key-value cache memory requirements for large language model inference by at least 6 times, and increase attention computation speed up to 8 times on H100 GPUs, while maintaining zero precision loss. This breakthrough has the potential to reduce AI deployment costs and accelerate the development of long context applications.

The rise of autonomous agents like OpenClaw has driven AI from conversation to task execution. Enterprise applications face challenges such as computing power waste and security compliance, with large-scale deployment becoming critical. Zhang Peng from Ant Digital pointed out that this marks a new phase in enterprise-level AI applications.

Chinese AI company Moonshot Ventures is considering an IPO in Hong Kong to take advantage of the market's enthusiasm for AI and strengthen its position. The company has held preliminary discussions with China International Capital Corporation and Goldman Sachs, and the preparation for the listing is still ongoing.

AI transformation often faces three major pain points: high costs, instability, and difficulty in reimbursement. Aiyiwei provides enterprises with a unified API hub, integrating 534 global models, achieving unified management and standardized interfaces. Technical teams can easily compare model performance and switch between models at zero cost for A/B testing, solving the challenges of dealing with multiple vendors and managing them, making technical decisions more efficient.

The AI Yiwei platform aggregates 534 mainstream large models globally, such as GPT-4, Claude, Gemini, etc., providing developers with a unified API interface to solve the problems of switching between multiple platforms and complicated registration and recharge processes, improving development efficiency and achieving AI freedom.

Artificial intelligence has not yet triggered mass unemployment, and the labor market remains generally healthy. Data shows that the unemployment rate in high-AI-dependent industries is not significantly different from that in low-exposure industries. However, AI may impact entry-level white-collar jobs in the next five years, requiring the establishment of corresponding coping mechanisms.

Chinese AI company Moonshot is considering an IPO in Hong Kong and has already conducted early assessments with investment banks such as CICC and Goldman Sachs. Riding the popularity of its Kimi large model, the company has joined the ranks of "large model unicorns" competing for listings, adding further momentum to the AI investment boom.

Elon Musk's AI firm xAI faces major staff turnover, with only one of its 11 co-founders remaining after recent departures, raising concerns over management and strategy.....

Ant Group and Shanghai Jiao Tong University jointly released the F2LLM-v2 series of Embedding models, aiming to break the English-centric bias in the field of semantic representation. The model swept 11 SOTA rankings in the authoritative MTEB benchmark, demonstrating dominant performance. As a fully open-source solution, it combines high performance with extreme efficiency, providing global developers with advanced semantic representation tools.

Microsoft OneDrive launches an 'AI Reshaping' tool to address the issue of distortion caused by excessive beautification in AI-generated images. This feature can accurately retain the core subject of a photo while changing the artistic style, avoiding image drift and achieving diverse styles while keeping the subject unchanged.

Apple and the University of Wisconsin-Madison jointly introduced the RubiCap AI training framework, focusing on dense image description, aiming to enable AI to accurately describe image details such as a red apple on a table. The framework uses reinforcement learning, achieving more with less, leveraging Qwen2.5 as a referee to improve training effectiveness.

ByteDance open sources the super intelligent agent orchestration framework DeerFlow 2.0, which has gained 44k stars on GitHub and topped the trending list. The framework is deeply adapted to local scenarios and is called by developers as "OpenClaw更适合中国宝宝" (more suitable for Chinese babies). Its core is multi-dimensional capability integration, not just a conversation tool, but also a SuperAgent scheduling center, supporting sub-agent orchestration and enabling the construction of complex task processes.

Tengyun Digital Technology launched the film-level AI long video creation platform CineART, marking the entry of AI video creation into the 'targeted production' era. The platform is based on its self-developed large model and restructures the creative workflow with 'director's logic,' aiming to solve the core issues in current AI video generation and achieve a leap from prompt generation to full-process professional creation.

DingTalk launched Wukong AI, focusing on low-barrier deployment, allowing enterprise users to easily use it without a professional team. The product is positioned as a corporate-level AI assistant, emphasizing security, ease of use, and personalized experience, aiming to simplify the application of AI technology and redefine the standard for office AI.

Cloudflare introduced the open-source model Kimi K2.5 from Moonshot AI, aiming to improve the cost-performance of computing power. This model features an ultra-large context window of 256k, excellent visual processing capabilities, and stable multi-turn tool calling performance, showing outstanding results in programming and Agent tasks. It has passed rigorous evaluation and is now applied to core business operations.

Google introduced TurboQuant technology, which effectively addresses the memory bottleneck in large language model inference by compressing the KV cache. It significantly reduces memory usage without compromising accuracy, improving efficiency for processing long texts and complex tasks.

The MIIT has publicly solicited opinions on 121 industry standard plans, focusing on regulating the application security of artificial intelligence model context protocols. The goal is to address protocol compatibility and data security issues in large models related to multimodal interaction, long text processing, and cross-platform calling through standardization, marking a significant step forward in China's AI underlying protocol standardization and security regulation system construction.

ByteDance's CapCut recently launched two AI features: Video Studio and AI Video, integrated with the Dreamina Seedance 2.0 model, simplifying the video creation process. Video Studio replaces the traditional timeline with an infinite canvas, enabling non-linear editing; AI Video supports text-to-video generation, improving creative efficiency.

GitHub announced that starting April 24, 2026, it will use user interaction data to train AI models, causing controversy. This move has been jokingly referred to as a 'revert' operation, as GitHub had previously emphasized respect for user privacy. The new policy is opt-out by default, meaning private repository data will no longer be absolutely private, and even the free version of Copilot will be affected.