AI 今日新闻 · 2026-05-18

Shanghai Jiao Tong University x ShangHai Creation x Rui Jin Hospital Unveil CX-Mind: Chest X-ray Diagnosis Enters the Era of 'Verifiable Reasoning'，through multimodal large models and reinforcement learning technology, enhancing the explainability and clinical utility of medical imaging AI.

A Representation Entanglement Framework for Model-Heterogeneous Federated Learning

Thoth is a scientific reasoning model for bio-experimental protocol generation, using structured component-based reward mechanisms to ensure executable, evaluable, and interpretable protocols.

Huawei drives a body-aware intelligence startup boom through internal talent flow and external policy support, forming a comprehensive industry ecosystem.

This article discusses the transformation path of embodied intelligence in industry and business, analyzing the current status, challenges, and future of four furniture companies (ABB, Topsda, Yutu, Xingdongjinyuan), revealing key issues in the industry.

AlphaGo uses MCTS and neural networks to achieve efficient search, showcasing the potential of reinforcement learning.

Tasklet reengineers its technology stack to build a neutral multi-model platform, with future software companies divided into three categories: horizontal platforms, headless APIs, and solution companies.

NVIDIA GEAR Lab introduces DreamDojo and DreamZero world models, advancing embodied intelligence.

Greg Brockman highlights how Codex can be used to build applications on your phone.

用一段话概括文章最重要的信息，适合直接显示在资讯卡片上。

OpenAI Developers on X: The local server list got a cleanup pass. Better filtering, remembered sort state, clearer empty states, connected route state, and refreshes for non-listed ports every 120 seconds.

Abridge is revolutionizing healthcare by integrating AI into clinical documentation, saving clinicians 10–20 hours weekly, and enabling real-time prior authorization.

This course dives deep into JavaScript's internal mechanics to help developers build engineering-grade thinking models.

vLLM model has a serious 'token swallowing' issue under high-concurrency scenarios, which has been fixed by FanShi team.

AI is transforming traditional industries into new technological bottlenecks, with memory as the latest example.

Cursor AI's Agent request volume surged 15 times, 75% of enterprise code is generated by AI, 30% of PR is completed end-to-end by Agent, exploring fully autonomous team model without human involvement.

The article discusses new trends in the AI field, emphasizing the importance of recursive AI architecture and highlights Tian Yuandong's collaboration with top teams to drive AI evolution.

Kimi WebBridge enables AI agents to operate browsers like humans, challenging Codex's closed ecosystem.

The article points out that LLM has inherent defects, open source will catch up with closed source, and predicts that OpenAI may become the next Sun company.

Simon tracks the name evolution of the OpenClaw project, revealing its technical development path.

Vercel AI Gateway now supports sorting model providers by cost, latency, or throughput to help developers choose services more efficiently.

Wanderingmeow's detailed feedback and confirmation of project functionality demonstrate the value of non-code contributions to open source.

Vercel CLI now supports using curl syntax to call deployed APIs.

Meituan opensource General 365 benchmark, revealing the real capability boundary of large models in general reasoning.

Companies still need real people responsible for critical work, AI cannot completely replace human responsibility.

Those who master agent management and fundamental principles will be unstoppable, as agents significantly amplify output.

The article emphasizes the critical nature of the AI competition in 2028, highlighting the need for the US to maintain leadership in chips, talent, and export controls to avoid being overtaken by China.

Vercel CLI now supports native curl commands, allowing developers to test all deployments with Vercel authentication.

Grok CLI through plugin and skill support achieves cloud deployment enhancement, and the Vercel plugin provides seamless deployment capabilities.

The article points out that the ideal state does not exist in reality and emphasizes that technical practice should be based on real needs.

Paul Graham emphasizes that investors have no right to force a company to be acquired if the founder does not want it.

Xiaomi air conditioner achieves intelligent control and energy efficiency optimization through AI engineering practices, enhancing user experience.

Paul Graham emphasizes that well-informed optimism is more powerful than blind optimism and must be based on facts and analysis.

The article introduces how to use AI agents combined with traditional development tools to solve the problem of unstable tests by comparing different execution paths to find the root cause.

The top 1% of AI users use the RSTI prompting framework to improve results.

Paul Graham points out that lawmakers should not just park their money in index funds or blind trusts, but should actively trade to protect their families.

This article summarizes the latest updates and trends in the Java and Kotlin ecosystems for May 2026, including AI tool usage, test-driven development practices, and technical insights.

Combining Spring Data JDBC with Kotlin provides a simple, predictable persistence model suitable for projects requiring immutability and null safety.

Grok Build provides fully free unlimited APIs, making it the easiest way to use Grok Build CLI for 100% free.

The article focuses on the latest developments in AI and technology, including Apple's Siri update, Malta's ChatGPT subsidy, Xiaomi's battery upgrade service, and SpaceX's IPO plan.

The article introduces the latest AI model released by OpenAI, highlighting its breakthroughs in natural language processing and code generation.

Mounting evidence of productivity lift across high-AI industries.

Arena.ai uses a unified tagging system to process millions of votes per week, with a data pipeline built on Databricks and Spark.

Anthropic's Claude Code now supports operation on small devices, marking progress in edge computing.

The hidden cost of AI agents may be much higher than expected, and using 1000 times more tokens may not necessarily lead to better results.

The reconstruction of AlphaGo highlights key insights into self-play, reinforcement learning, and the future of large language models.

A world model is a significant advancement in reinforcement learning that combines environment modeling with decision-making, becoming a billion-dollar investment hotspot in AI.

Firecrawl v2.10 introduces local file parsing, a fully sandboxed scrape mode, three new scrape formats, and four new SDKs, along with numerous reliability fixes.

OpenAI fully covers the token bill of the 'Lobster Father', which cost $13 million in a month.

Monzo builds a governed data mesh across 100 teams and 12,000 dbt models.

The sudden emergence of art 50,000 years ago may not be due to a single genetic mutation but rather polygenic adaptation.

Yao Teacher's WeChat Reading Skill generates visual reports to help users analyze their reading data.

The video introduces an experiment from the Fireship channel, demonstrating how to achieve an astonishing feature through code and tools.

OpenAI has open-sourced Symphony, a SPEC.md for autonomous coding agent orchestration.

By configuring Hermes, FeiShu robots can use different models to achieve batch @ multiple members in a group.

Feishu robots default do not support mutual @, but a custom skill can achieve stable mutual @.

The OpenAI Codex team is observing how engineering organizations are changing as AI agents become part of the team, with a live discussion event scheduled for May 21.

Vercel integrated Stripe to enable developers to monetize apps seamlessly, becoming the third most popular integration on the platform.

OpenAI is building a device network based on Codex, enabling multi-device collaboration.

This article explains the three core areas of AI engineering: prompt engineering, context engineering, and harness engineering, providing practical guidance for product managers.