AI 今日新闻 · 2026-05-22

本文介绍了如何通过Auth Proxy来保护LangSmith代理沙箱的网络访问，确保在大规模部署代理时的安全性。Auth Proxy通过在网络层控制和管理代理与外部服务的交互，实现了凭据的安全管理、网络访问的显式控制以及团队职责的清晰分离。

VSCode团队提出Agent-First Development五大支柱：模型选择、行动边界、上下文、提示精度和工具控制，强调从人+编辑器转向人+Agent+编辑器的开发范式，通过精细化配置提升AI编程效率。

Text Arena数据显示自2023年以来AI模型价格性能比发生巨大变化：GPT-4级别质量成本降低500倍，从每百万token约50美元降至0.10美元，低端模型性能大幅提升而高端模型价格下降。

CSS新增sibling-index()和sibling-count()函数，让开发者无需JavaScript或复杂的:nth-child规则即可实现动态元素索引计算，一行代码解决级联动画延迟问题，支持任意数量元素。

微软研究团队推出Vega零知识证明系统，可在不暴露凭证本身的情况下验证政府颁发的身份信息，支持移动端100毫秒内生成证明，无需可信设置，即将开源。

微软研究院发布MagenticLite、MagenticBrain和Fara1.5三个组件，专为小型模型优化的智能体体验，通过协同设计实现浏览器和本地文件系统统一工作流，其中Fara1.5在网页导航性能上几乎翻倍提升。

DeepSeek 的视觉指针机制让开源模型用 90% 更少视觉 token 在 7 项基准追平或超越 GPT-4V，同时提供可回溯的可解释推理路径。

OpenAI未公开的通用推理模型自主解决Erdős 1946年提出的平面单位距离问题，通过代数数论工具解决离散几何问题，证明强推理能力达到阈值后创造性会自然涌现。

本期科技爱好者周刊聚焦于AI技术的快速发展及其对社会财富分配的影响。文章指出，AI相关产业如内存、储存、CPU、服务器等的股价大幅上涨，表明财富正迅速向AI领域集中。此外，文章还探讨了AI在日常生活中的应用，如通过AI估算食物的碳水含量，但实验表明AI在这方面并不准确。微软宣布将淘汰短信验证码，转而采用更安全的验证方式，如Passkey。亚马逊推出供应链服务，开放其物流网络，可能对制造业产生影响。最后，文章介绍了一个机械打字机模型玩具，以及几篇技术文章，包括关于GitHub Pages域名盗用问题、JavaScript ShadowRealm API和Firefox配置指南。

阿里巴巴推出全新升级的超大规模语言模型 Qwen3.7-Max，该模型专为代理中心工作设计，如编码、办公和生产任务以及长期自主执行。相较于前代 Qwen3.6，Qwen3.7-Max 在编码和代理基准测试中取得了显著进步，并引入了显式提示缓存功能，以优化重复上下文的处理。

阿里巴巴推出Qwen3.7-Max，作为面向代理时代的最新旗舰模型，它是一个多功能的基础模型，适用于能够实际完成任务的代理。该模型在编码代理方面表现出色，能够进行前端原型设计、多文件重构和实际调试。此外，它还是一个可靠的办公和生产力助手。

本文介绍了如何通过显式缓存优化Qwen模型的使用，包括缓存的工作原理、实现方法和最佳实践，帮助用户提高效率并降低成本。

OpenAI 的模型在解决平面单位距离问题上取得了突破，这一问题自1946年首次提出以来一直困扰着数学界。传统的观点认为最佳解决方案类似于方形网格，但 OpenAI 的模型发现了更优的配置，挑战了这一长期存在的假设。

Perplexity has implemented query-aware compression in their search system, which reduces context tokens by up to 70% while improving answer quality, leading to faster, cleaner, and more accurate search results.

SpaceX在IPO文件中列出了Grok的“辛辣”模式作为风险因素，这可能会影响其卫星互联网服务Starlink的运营。

本文介绍了 LangChain 的沙盒 Auth Proxy，这是一种控制代理生成行为与外部世界之间边界的工具。通过使用 Auth Proxy，可以安全地管理代理对网络资源的访问，防止未授权的访问和潜在的安全风险。

LangChain introduces a new streaming protocol for agents, aiming to make building applications feel more like developing software and less like parsing logs. The protocol provides typed projections that apps can subscribe to, addressing the limitations of token deltas in real-world applications.

构建智能代理的最艰难事实是，只有在生产环境中才能真正了解它们的行为。LangChain 的联合创始人 Harrison Chase 强调了在开发和部署智能代理时面临的挑战，包括不可预测的行为、安全性和责任问题。他建议通过在受控环境中进行测试和监控来减轻这些风险，并强调了持续学习和适应的重要性。

AI agents can now write production code, call external APIs, and manage pipelines autonomously, but the access layer lags behind, assuming human intervention for credential provisioning. Mem0's Agent-First mode addresses this by enabling agents to set up credentials without human oversight, streamlining the process.

Mem0 introduces Agent-First, enabling AI agents to self-provision memory in under 5 seconds without human intervention. This feature allows agents to onboard themselves and contribute to a shared memory project in the AGENTRUSH competition, which aims to assess the quality of memories created by agents through peer retrieval.

Climate tech companies are shifting their focus from decarbonization to critical minerals and supply chains due to political and financial pressures. Boston Metal, for example, is pivoting from low-emission steel production to producing critical metals like niobium and tantalum, which are in high demand for various industries. This shift is seen as a way to secure funding and stay afloat in a challenging environment, potentially paving the way for future climate benefits.

Naval Ravikant在本期播客中分享了他对销售的独特见解，他认为销售的本质不是说服，而是将真相清晰地传达给对方。他强调了可信度、诚实和理性共情在销售中的重要性，并提出了在交易中关注上行空间和长期主义的观点。

Daytona's Agent-Native Compute platform is designed for AI agents, offering ultra-fast sandboxes, high startup rates, and massive daily runs, making it ideal for reinforcement learning and evaluations. The platform has pivoted from human developer environments to focus on agent sandboxes, emphasizing bare metal performance and stateful snapshots. With RL workloads accounting for nearly half of its usage, Daytona is redefining the AI cloud landscape, potentially shifting it towards a model similar to Stripe rather than AWS.

本文介绍了Tiny LLMs和Agents在边缘设备上的应用，特别是Function Gemma模型在Pixel 7上的性能表现，以及开发者在设备上实现AI的两种路径：基于Gemma 4的技能框架和Eloquent生产转录应用。

In the era of AI, scaling creativity is essential for meeting the insatiable demand for fresh, unique media. AI can help by absorbing repetitive tasks, allowing creative teams to focus on strategic decisions. However, responsible adoption is crucial to maintain brand integrity and build customer trust. Companies like Nestlé are already leveraging AI to generate brand-informed assets and improve workflow efficiency.

尽管传统RAG在处理代理工作负载时存在局限性，但通过引入代理RAG，可以有效解决这些问题。代理RAG通过查询路由、混合检索、检索评估和多步检索等机制，使得检索层与工作负载相匹配，从而提高系统的性能和可靠性。

Zilliz Cloud maintains fast and accurate filtered vector search at scale through two strategies: preserving graph connectivity during filtering and switching to brute-force scans for highly selective filters.

The article discusses the need for extraordinary government intervention in response to AI risks, arguing that while AI's economic impacts are manageable, its misuse risks require significant action. The author suggests that improving societal resilience is a better approach than restrictive measures on AI development and deployment.

Google AI 发布 Gemini API 的 Managed Agents Quickstart，提供预构建的代理，帮助开发者快速构建和部署智能代理，无需从头开始。这些代理包括搜索助手、代码助手和文档助手，基于 Gemini 模型，具有强大的语言理解和执行任务能力。

Philipp Schmid展示了如何使用单个curl命令调用Gemini API来构建一个GitHub问题分类代理，该代理能够克隆仓库、抓取开放问题、分类问题类型并执行复现代码以确认bug，整个过程无需复杂的编排框架或基础设施。

In this interview, Ben Thompson speaks with Parag Agarwal, the founder of Parallel, about the future of content valuation and creation incentives in an era dominated by artificial intelligence and autonomous agents. They discuss how the advent of AI and the concept of the 'agentic web' are reshaping the way content is valued and created, and explore potential solutions for sustaining high-quality content in this new landscape.

Large language models (LLMs) can replicate average responses of major household surveys, but they fail to capture the dispersion of responses, leading to a 'mode collapse' where the model's responses are too homogeneous. The paper 'Can LLMs Mimic Household Surveys?' explores this issue and attempts to address it through unlearning techniques, showing some improvement in capturing the variability of human responses.

Weaviate v1.37.1 introduces an MCP server integrated into the database, enabling efficient codebase ingestion and hybrid search for coding assistants like Claude Code, Cursor, or VS Code. This feature addresses context window limitations and improves code query handling.

In 2026, data scientists need to master three key skills with Claude: data analysis, model evaluation, and ethical considerations. These skills are essential for leveraging Claude's capabilities effectively in the field of data science.

NVIDIA has introduced NVIDIA-Verified Agent Skills, which enhance AI agents' capabilities while ensuring transparency and security. These skills provide detailed information about their functionality, origin, associated risks, and any modifications, adhering to the agentskills.io open specification for compatibility across various AI platforms.

This article presents 10 GitHub repositories that are essential for mastering quantitative trading. These repositories cover a wide range of topics from basic trading strategies and frameworks to advanced portfolio optimization and machine learning approaches. They are valuable resources for both beginners and experienced traders looking to enhance their skills and knowledge in quantitative trading.

This article explores advanced SQL window functions beyond basics, focusing on solving real business problems through four key patterns: running totals, gaps and islands (sessionization), ranking and classification, and data imputation. It provides practical examples and code snippets to illustrate how window functions can be effectively utilized in data analysis and manipulation.

Qwen3.7-Max是阿里云推出的新一代旗舰级AI模型，专为代理时代设计，具备多功能基础，能够实际完成任务。它在编码代理、办公和生产力辅助方面表现出色，支持长周期自主操作，并且与各种开发栈兼容。

This article demonstrates how to anonymize sensitive production data for data science using the Mimesis library in Python. It provides a step-by-step guide to install Mimesis, generate synthetic data for personal identifiable information (PII), and replace real PII in a dataset with anonymized data, ensuring data privacy and compliance.

Qwen3.7-Max在编码代理和通用代理的基准测试中表现出色，尤其在最难的推理基准上表现出色，并在通用能力和多语言支持方面脱颖而出。

Qwen在自主执行过程中，通过连续运行约35小时，进行了1158次工具调用，完成了432次内核评估，自主编写、编译、分析和迭代改进了Extend Attention Kernel，实现了10.0倍的几何提升。

This article highlights the advancements in small language models, specifically those with under 7 billion parameters, which can now run on consumer GPUs or even laptops. It emphasizes that these models are now capable of performing tasks that were previously only achievable by much larger models, thanks to improvements in training data quality, distillation techniques, and architectural innovations like Mixture-of-Experts (MoE). The article provides a curated list of the best small language models available on Hugging Face, along with their capabilities and benchmark scores.

Qwen3.7-Max 在人工智能分析指数上获得了56.6分，比Qwen3.6-Max-Preview提高了4.8分。它在科学推理、代理能力、编码能力和减少幻觉方面都有显著提升。

Thomas Wolf, a prominent figure in the field of natural language processing, has recently shared an interesting discovery about DNA modeling being distinct from language modeling. In an interactive blog post and demo, he explores this difference in depth, highlighting the unique challenges and opportunities presented by DNA sequences. This work is a collaborative effort between the Hugging Science, pre-training, and post-training teams, showcasing advancements in computational biology and AI.

微软因成本问题取消了内部Claude Code许可证，Uber CTO警告公司已耗尽2026年AI预算，这表明开源模型的使用正在加速增长，即将迎来新的转折点。

智谱发布GLM-5.1高速版API，实现400 tokens/s的全球最快大模型API速度，同时保持旗舰级能力，适用于AI编程、实时交互等高延迟要求场景。

Andrew Ng announces a new short course on building AI agents for generating images and videos, emphasizing the importance of self-evaluation and iteration for improving output quality. The course, developed in collaboration with Google Cloud, is taught by Katie Nguyen and Wafae Bakkali and focuses on three evaluation techniques: image-text similarity scoring, LLM judging against custom criteria, and structured rubrics for detailed assessment.

AI生成的代码导致40%的严重缺陷率和70%的总体缺陷率增加，大规模部署AI代码审查系统需通过实时评估优化流程，将代码审查作为主要开发瓶颈。

LLMs生成的代码在企业级应用中面临质量差距，需通过改进开发流程和工具来解决，以实现可持续的生产级代码生成。

Datasette Agent是首个结合LLM与Datasette的AI助手，支持通过对话查询数据并生成图表，基于Gemini 3.1 Flash-Lite模型运行，提供插件扩展能力。

FTC要求Cox Media Group等三家公司支付近100万美元，因其虚假宣传‘主动倾听’AI营销服务，实际并未使用语音数据，仅转售数据列表。

Vlad Feinberg的指南指出，掌握LLM内核级调优和MoE架构优化是进入前沿实验室的关键，同时Agent自动化和可观测性成为基础设施新趋势。

全栈代理和生成式UI正在推动AI交互的范式转变，AG UI协议作为代理用户交互标准已被Google、Microsoft等广泛采用，标志着AI界面从MS-DOS式黑屏向图形化时代过渡。

Railway通过自建裸金属数据中心和云突发策略实现3个月回本周期，以70%利润率支持代理原生云平台，35人团队服务300万用户并每周新增10万用户。

OpenAI的GPT-next模型以不足1000美元的成本，在32小时内解决了持续80年的Erdős平面单位距离问题，证明了通用LLM在复杂科学推理中的潜力。

CUDA-Q是一个统一的量子计算平台，整合GPU、CPU和QPU，解决量子计算中的算法、硬件噪声和错误校正挑战，支持跨硬件的量子工作流开发。

Daytona通过提供可组合、状态化的沙盒环境，解决了AI代理对动态计算资源的需求，其技术架构支持从零到10万CPU的弹性扩展，并成为AI基础设施的关键组件。

微软研究院提出Intervene框架，通过LLM-based projection将AI代理输出分解为可验证属性，并实时生成形式化规范以确保合规性。

OpenAI利用AI技术推翻数学界80年未解猜想，获菲尔兹奖得主认可，展示AI在数学研究中的新潜力。

OpenAI推出Codex内置浏览器的高级标注模式，支持直接调整页面元素、即时预览和批量评论，显著提升设计与开发迭代效率。