AI 今日新闻 · 2026-05-13

This article introduces a method for building long-running AI agents using ADK that maintain context even after pausing and resuming.

The attack surface of AI coding agents has expanded beyond source code to four categories of files: execution, instruction, connection, and extension, with Google Threat Intelligence using VirusTotal Code Insight for semantic-level threat analysis to effectively defend against supply-chain attacks.

k6 2.0 released with AI-assisted testing workflows, introducing 4 new CLI commands for deep integration with AI tools like Claude Code, boosting test automation efficiency by over 50%.

文章探讨了AI支出不再仅仅是IT预算问题，而是涉及整个组织的资本分配纪律。CFO们现在要求AI投资产生可衡量的业务成果，而非单纯的实验。

OpenAI和Google分别推出了新的部署公司，通过雇佣大量工程师帮助客户采用AI技术，这标志着AI行业正在回归到20世纪70年代的计算模式，即由企业高管主导的技术变革。

本文提出了一种针对生产环境中AI代理的12指标评估框架，涵盖检索、生成、代理行为及生产性能四个维度，帮助团队在部署前全面评估AI系统的可靠性与表现。

本文探讨了如何通过监督微调将一个小型语言模型转变为具有特定人格的C-3PO。实验表明，第一人称陈述比对话演示更有效，而合成文档则在传授角色知识方面表现更好。

本文探讨了美国在关键矿物供应和电力基础设施现代化方面的挑战，强调自动化、强化学习和垂直整合操作的重要性，以及如何通过技术创新和政策改革加速这些领域的进展。

Luke Alvoeiro 在播客中介绍了五种多智能体架构类型，强调了多智能体系统在解决注意力瓶颈方面的潜力，并展示了如何通过 Missions 系统实现长时间自主任务。

3Blue1Brown 解释了子集和问题，这是一个经典的计算机科学难题，探讨了其背后的数学原理和算法复杂性。

AI accelerates code production, but it does not change how quickly a codebase can safely absorb these changes. Ignoring this gap can lead to unexpected failures in production. The article emphasizes the importance of continuous refactoring to improve the system's ability to absorb frequent changes.

Google introduced Gemini Intelligence at the Android 17 launch, transforming Android from an operating system to an intelligent system with multi-modal and cross-environment AI capabilities, including features like auto-fill, smart voice input, and multi-step task handling.

Kaiming He's team has introduced the first continuous diffusion language model ELF, which achieves a perplexity of 24 on OpenWebText with only 105M parameters and 45B training tokens, outperforming mainstream diffusion models.

MatterSim expands AI's role in materials science, from faster large-scale simulations to MatterSim-MT, a new multi-task model that simulates properties beyond potential energy surfaces.

Augment Code used Mercury 2 as a dedicated subagent, achieving an 82% faster context compaction and 90% lower summarization costs.

Symphony uses Codex agents to automatically handle tasks, significantly improving development efficiency.

dnsmasq has released six serious security vulnerabilities' CVEs, affecting almost all non-ancient versions, with patches already available.

Karpathy points out that 90% of the AI coding bill is spent on unnecessary context, and optimizing context usage and routing strategies can significantly reduce costs.

The rise in memory prices is affecting multiple industries, from smartphones and computers to the automotive sector, even impacting the Korean dating market.

Be cautious when migrating bundling tools; unless the current solution severely impacts development, avoid blindly pursuing new technologies.

95% of GenAI pilots fail to reach production due to the 'myth of the last mile', while reinforcement learning (RL) can systematically improve models through continuous feedback and refinement.

The Google Cloud AI team emphasized that the user experience should always focus on the goal, while ensuring users can intervene to build trust when designing Gemini Enterprise.

The video explains three common AI terms: hallucination, context window, and tokens, helping to understand their actual meanings and use cases.

Amazon's mandate requiring 80% developer AI usage with Token tracking has led employees to artificially inflate metrics using MeshClaw. This tokenmaxxing phenomenon reflects Big Tech's ROI anxiety over $200B AI investments and exposes security risks in AI Agent tools.

Cloudflare discovered a bug in their QUIC implementation where the CUBIC congestion control algorithm failed to recover from network congestion, leading to connection failures.

You can now manage the Vercel Firewall directly from the CLI, including configuring custom rules, IP blocking, system bypass rules, attack mode, and system mitigations.

This article discusses how to build a Google Drive sync engine that runs smoothly within the strict limits of MV3 service workers, focusing on in-memory state management, offline handling, and replacing the Google SDK with native Fetch.

Docker's AI Governance solution ensures enterprise security by controlling agent execution paths and tool invocation paths, applicable to developers' laptops and other environments.

Databricks introduces native lakehouse sync, enabling seamless integration between data lakes and data warehouses, enhancing data sharing and governance capabilities.

Needle is a 26MB function call model that runs on incredibly small devices, suitable for edge computing and IoT applications.

LiteParse is an open-source, model-free document parser that supports over 50 document types, quickly parses complex text layouts and tables, and extracts clean text in seconds, with lightweight OCR integrations.

Agent Harness is the key engineering system that determines whether an AI model can truly execute tasks, including tools, permissions, context, execution environment, and result verification.

GB 200s improve the prefill and decode disaggregation efficiency for large MoE models like Qwen, significantly enhancing throughput compared to the Hopper platform.

Perplexity published new research on deploying the Qwen3 235B model on NVIDIA GB200 NVL72 Blackwell racks, showing that GB200 outperforms Hopper in high-throughput inference for large MoE models.

NVLS all-reduce latency significantly improves from 586.1µs on H200 to 313.3µs on GB200, with notable performance gains in MoE prefill and decode throughput.

NVIDIA platform, through various optimization techniques, becomes the best platform for large-scale model inference, significantly reducing service costs and improving performance.

Databricks Data Intelligence Platform helps sports teams leverage massive amounts of tracking data to improve player health, game performance, and overall operational efficiency.

World models are a significant trend in current AI research, focusing on how AI can better understand the real world.

AutoScout24 significantly improved development efficiency and code quality by adopting AI tools like ChatGPT and Codex, reducing project delivery times.

The Parameter Golf challenge demonstrated the potential of AI-assisted research, including training optimization, quantization, test-time and evaluation strategies, and new modeling ideas.

NVIDIA engineers and researchers use Codex for complex engineering work and end-to-end machine learning experiments, significantly improving development efficiency and system performance.

Despite the decrease in individual AI token prices, overall AI expenditure in enterprises is increasing due to more work being done through AI, leading to higher total costs.

Codex helps finance teams prepare monthly business reviews, reports, variance analysis, and planning more efficiently, without coding.

Temporal Reasoning layer manages memories with time signatures, significantly improving the accuracy of long-running agents' memories.

China's open AI ecosystem reduces redundant R&D compute costs, enhancing model development efficiency and sustainability.

Using GitHub Copilot CLI to build a procedurally generated roguelike game, showcasing practical applications of AI-assisted development.

This article provides a detailed guide on how to write, test, and deploy your first WebAssembly program and web app entirely within the browser, without installing any software.

Qdrant 1.18 is released, introducing TurboQuant, a new quantization method developed by Google Research, offering higher memory efficiency and better recall.

Proxy-Pointer framework through structure-aware enterprise document intelligence achieves efficient and precise document comparison, suitable for complex financial agreements and academic papers.

The article discusses hybrid search and re-ranking techniques in production RAG systems, addressing the limitations of dense vector retrieval in specific technical queries.

Jina AI releases the v5-omni model, retaining the v5-text backbone and adding pretrained vision and audio encoders connected through small trainable projectors.

This article explores the necessity and practical methods of transitioning from vibe coding to spec-driven development, emphasizing the advantages of the latter in team collaboration and project management.

Senior engineers can expand their influence and drive change by focusing on technical clarity, team alignment, and intentional documentation.

In production, demoting results works better than removing them because the latter excludes results that specific users need.

This article introduces five practical Python scripts for handling common tasks in time series data, including resampling, anomaly detection, and trend decomposition.

Polars outperforms Pandas in handling large datasets, especially in parallel computing and lazy evaluation.

MatterSim-v1 makes progress in experimental validation, faster simulation, and multi-task models, improving material design efficiency.

A multi-agent RAG system may produce errors due to retrieving low-relevance or stale documents, yet the output appears confident and correct.

This article introduces seven leading LLM observability tools that help AI engineers monitor, evaluate, and debug large language model applications running in production.

AWS proposes a hybrid multi-tenant architecture that addresses isolation and efficiency issues in large-scale ad-serving platforms through pre-integration models, Route 53 weighted routing, PrivateLink connectivity, and tiered architecture.