AI 今日新闻 · 2026-05-25

The LingBot-VA model developed by Ant Group and HKUST was accepted by RSS 2026, enabling robots to reason and act in real-time.

GPT 5.5 and other models' capability improvements are not sudden jumps but result of model reliability crossing a key threshold. Reinforcement learning, post-training optimization, and evolving evaluation systems drive AI practicality.

Google invested $40 billion in Anthropic's secret AI model, signaling strategic importance.

A Chinese AI company has broken the bottleneck of running a 60 billion parameter model on mobile devices using ternary quantization, saving 6x memory with minimal performance loss.

SaaS-Bench evaluation shows mainstream large models have less than 4% complete pass rate on real office tasks, revealing huge challenges for AI fully automated office work.

The open-sourced version of LongCat-Video-Avatar 1.5 significantly improves lip-sync accuracy, physical realism, and long-video stability for commercial-grade digital human video generation.

DeepSeek builds a low-cost, high-efficiency model system through multiple foundational innovations to drive China's $10 trillion AI hardware ecosystem and achieve its own $1 trillion valuation.

DeepSeek reduces KV cache requirements through innovations, driving China's AI hardware ecosystem toward a $10 trillion industry.

Shen Yujun believes that VLA and world models are not the endgame for embodied intelligence; a unique physical-world model will emerge. Robots lack data, so spatial perception must be built from sensor inputs.

Azure NetApp Files addresses EDA workload scalability issues in the cloud through high-concurrency, low-latency storage support.

AI won't replace human jobs but instead create new roles, especially those involving AI collaboration.

Future AI compute power will be dominated by inference at 70%, with communication energy consumption being the main bottleneck in data centers, and physical AI development depends on high-quality real-world data.

Hyper3D 发布 Rodin Gen-2.5，全球首个千万面级3D生成模型，4秒生成百万面模型，细节还原至毛孔级别。

DeepSeek 通过永久降价和优化技术，降低了大模型 API 的成本，使其更具性价比，有望吸引更多开发者和企业用户，从而挑战海外头部模型的地位。

Dan Shipper 预测未来一年 AI 自动化将带来更多需要人类判断和管理的工作，提出公司内部超级 Agent 和 AI 工作操作系统两大趋势。

小米对空调安装'抽真空'造假行为的处理，以及对抽真空环节重要性的探讨。

AI won't eliminate jobs but instead creates new roles, especially for PMs and designers.

Future AI inference will consume 70% of computing power, communication energy consumption can reach hundreds of times that of computation, and high-quality data and edge computing will become key bottlenecks.

JunoBrain proposes a cognitive world model based on cognitive science to achieve higher-level embodied intelligence and has secured billion-level funding.

AI agents do not reduce engineers' need to understand systems; instead, they elevate requirements for judgment and foundational knowledge.

Harness Engineering通过AI主导的组织架构实现了极高的开发效率，99%的代码由AI完成，每天平均3到8次生产部署，强调信任AI而非仅将其作为工具。

Humans should merge with AI to survive existential challenges, just as cells once merged with mitochondria to evolve complex life.

Reasonix is a terminal-native AI coding assistant designed for DeepSeek API, achieving up to 94% cache hit rate and 2.5x cost reduction via byte-stable caching.

Agent products are transitioning from concepts to real applications, with a significant surge in 2026. C-end users are declining while B-end and developer users are growing. Product forms are stabilizing, and iteration speed is accelerating.

Anthropic's Mythos 1, Claude Opus 4.8, and GPT-5.6 are rumored to be in development, with Mythos 1 showing promising results in finding vulnerabilities in open-source projects.

FaceWall Intelligence and partners have released BitCPM-CANN, the first 1.58-bit edge-side large model trained entirely on Huawei Ascend in China, achieving up to 6x memory savings while maintaining over 90% performance.

Claude Design 团队通过高频用户对话、极速迭代和自建工具，成功将设计工具从概念推向市场，实现了从自然语言到生产代码的全链路打通。

DeepSeek V4 price drops significantly, with Reasonix tool reducing costs to 20%, achieving a 99.82% cache hit rate.

The TrapDoor supply chain attack uses AI assistant configuration files as an attack vector, stealing developer credentials via malicious packages and PR injection.

Google Cloud reports a critical ViewState deserialization vulnerability in KnowledgeDeliver that allowed attackers to achieve remote code execution using publicly shared ASP.NET machine keys.

Chinese-language phishing services are evolving from static password harvesting to real-time interception and tokenization, using RCS/iMessage to bypass traditional SMS filters, targeting unauthorized financial account control.

Waymo recalls vehicles and halts Robotaxi services due to flooding issues, exposing limitations in perception and rule-based systems; Tesla adapts via FSD learning human driving behaviors.

Migrating from Go to Rust focuses on correctness guarantees, runtime tradeoffs, and developer ergonomics rather than performance differences.

The robot industry is on the brink of a breakthrough, where AI and cloud infrastructure will reshape engineer growth and industrial ecosystems.

Google CEO Sundar Pichai admits Google lags in AI coding and agent capabilities but emphasizes that search won't be fully replaced by AI; AI will reshape work rather than simply replace it.

Andrej Karpathy joined Anthropic to lead AI-powered pre-training research using Claude, marking a strategic shift toward autonomous AI development.

Heterogeneous intelligence is the key paradigm for future AI development, enabling efficient collaboration among diverse architectures, scales, and hardware models.

DeepMind uses a self-built platform architecture to run agents at scale with an agent manager and internal toolchain for efficient automation.

AI assistants like Claude cannot replace real architects; their designs lack understanding of teams, constraints, and real-world environments, leading to blindly adopting impractical architectures.

Data scientists should master API and documentation skills to enhance collaboration, reproducibility, and business integration.

尽管存在“SaaSpocalypse”的说法，实际业务数据显示AI并未显著冲击SaaS市场，许多传统软件公司反而展现出更强的韧性。

Node.js 提议内置虚拟文件系统，引发关于AI生成贡献的辩论。

Gemma 4 introduces multi-token prediction technology, achieving up to 3x faster token generation, significantly improving large model inference efficiency.

Microsoft Research introduces the Intervene mechanism, converting AI agent behaviors into verifiable properties and generating Python validators, significantly improving small models' performance on complex tasks.

Nine-Arm Skills is a lightweight GitHub project that enhances AI coding agents via structured prompt engineering for debugging, postmortems, and communication.

Google I/O 2024 introduced the Gemini 3.5 Flash and Omni models, plus the Gemini Spark agent system, emphasizing multimodal capabilities and cloud-based automation.

Google introduces new AI search features including conversational mode, image/file uploads, and personal intelligence integration to enhance user experience.

JDK 27 introduces G1GC as default, Vector API, and Compact Object Headers to enhance performance and developer productivity.

About 50% of expert computation in MoE models is wasted on tokens that don't require expert processing; ZEDA can skip such computations to improve efficiency.

ZEDA is a novel MoE technique that uses self-distillation to dynamically skip experts, improving inference efficiency and giving models compute budget awareness.

Datasette Agent 0.1a4 integrates with a new JS plugin hook to enable launching an AI chat interface via shortcut key for enhanced data analysis interaction.

Datasette 1.0a30 introduces a customizable "Jump to" menu for improved navigation and plugin extensibility.

Mole 1.5.0 was developed in six days with major enhancements across ten functional modules, significantly improving system monitoring, software management, and uninstallation capabilities.

Introduces a local kanban tool that manages AI agents via cards and supports parallel execution of tasks.

Tau is a free multi-provider AI agent tool that supports 23 providers and offers full agent capabilities with simplified credential management.

YC's 2026 startup list reveals that AI entrepreneurship is shifting focus from productivity tools to industry restructuring with deep vertical integration.

Browser Use launches a Rust-based terminal UI and browser control framework supporting any LLM with native Chrome automation.

Datasette 1.0a30 introduces a "jump to" menu accessible via the "/" shortcut, enabling quick navigation to databases, tables, or queries, and supports plugin extensions.

Abinitio Bio develops foundational models for biomanufacturing, reducing pharmaceutical process decision time from 6–18 months to hours, potentially saving over $100 million monthly in delays.

The cost of AI computing power continues to decline, driving broader industry adoption.