AI 今日新闻 · 2026-05-20

Vibe Coding is merely the starting point of a software production revolution; the next stage is the software factory—a new engineering paradigm where multiple AI agents collaborate and validate outputs using correctness benchmarks, with memory and skills becoming the core units of collaboration.

Cerebras is now running the trillion-parameter Kimi K2.6 model in enterprise trials at ~1,000 tokens/s, shattering the old belief that open-source large models are impractical due to hardware limitations.

Vercel now offers all firewall mitigation features — including user-configured rules — completely free of charge, eliminating fees for blocked, challenged, or rate-limited traffic and reducing edge security costs significantly.

AI automation replacing human workers destroys its own customer base; tech elites promote UBI not out of charity but as a modern form of forced consumption to sustain subscription economies, mirroring 19th-century indentured labor systems after slavery abolition.

2026 will be the year of AI Harness. Using engineering methods like guardrails, validation, and automation processors, unreliable AI Agents can be transformed into stable, controllable systems without modifying Prompts, marking key infrastructure for AGI.

This article details a production-grade deployment of a multistage multimodal recommender system on Amazon EKS, achieving millisecond latency and real-time updates for millions of items using Bloom filters, in-memory feature caching, and Kubeflow-based continuous fine-tuning.

The top 10 Python libraries for data engineering in 2026 revolutionize pipeline construction across orchestration, ingestion, quality, and storage—especially Prefect, SQLMesh, dlt, and Bytewax, which drastically reduce operational complexity and boost maintainability.

In agentic AI systems, prompt engineering has evolved into context engineering: reliable behavior requires deliberate design of four components—system prompt, tools, examples, and context state management—otherwise, context rot causes behavioral drift across multi-step reasoning, as identified by Anthropic's research.

Google released Gemini 3.5 Flash at six times the price of its predecessor, yet deployed it across Search, AI Assistant, and enterprise tools—revealing a strategic shift toward internal model saturation over API monetization.

Pake v3.11.5 enables turning any webpage into a cross-platform desktop app with one command, built on Rust/Tauri with ~5MB size, fixing critical crashes on Linux Wayland and macOS while adding dock badges and incognito mode.

Meta proposes AIRA, a dual-agent system that autonomously discovers neural architectures outperforming Llama 3.2 at 350M, 1B, and 3B scales within a 24-hour compute budget, offering a reusable engineering paradigm for AI agent design.

Vercel launches a flat-rate CDN pricing model that eliminates overage fees for traffic spikes and viral events, delivering predictable costs without compromising network performance.

Cloudflare and Anthropic have launched an integration enabling developers to run Claude Managed Agents within Cloudflare Sandboxes, delivering secure, scalable AI agent deployment with private service access without exposing internal systems.

Tencent's Marvis breaks through the limits of chatbot-style AI by integrating at the OS level, enabling users to control files, system settings, and cross-device apps via natural language — the first true personal AI assistant that understands and acts on your computer.

Marc Brooker asserts that the commercial scale of Agentic AI is gated by defect rate, not model capability; only pushing both frequency and severity of defects into the “low-low” quadrant unlocks trillion-dollar knowledge-work markets.

eBPF provides security observability with kernel-level visibility and protection that user-space agents cannot match, as probes attached directly to the Linux kernel syscall interface remain functional even when attackers have container root, while reducing security-related CPU overhead by 60-80%.

Google Cloud announced Gemini 3.5 Flash and Gemini Omni models at I/O, alongside Gemini Spark agents and CodeMender security tools, significantly enhancing enterprise AI capabilities in video generation, coding, and automation.

Snapchat uses the Bento ML platform to retrieve, score, and rank content from millions of items within 100ms, supporting real-time recommendations for 477M daily active users.

Google releases the Gemini 3.5 family, starting with 3.5 Flash for complex agentic workflows. It outperforms 3.1 Pro on coding and agent benchmarks and runs 4x faster, reaching 12x in Antigravity.

Amazon Bedrock now supports Programmatic Tool Calling (PTC), allowing LLMs to generate Python code for batch tool execution instead of sequential round trips, dramatically reducing latency and token usage for multi-tool workflows. AWS offers three implementation paths: self-hosted Docker sandbox, Bedrock AgentCore Code Interpreter, and Anthropic SDK-compatible proxy.

Enterprise AI deployment faces pressures from regulation, sovereign clouds, and latency, forcing architectures to shift from cloud to on-premise or edge for compliance and real-time response.

LangChain introduces the Agent Development Lifecycle (ADLC), dividing agent development into four phases—build, test, deploy, and monitor—emphasizing that its fundamental difference from traditional software lies in infinite input/output spaces and non-determinism, with successful teams adopting a "ship early, iterate fast" pattern.

Claude Cowork is a desktop autonomous AI agent that completes complex knowledge work—like document generation and data organization—by directly accessing local files without manual intervention, dramatically boosting efficiency for non-technical professionals.

Databricks releases a real-time fraud detection solution based on Spark RTM and Lakebase, achieving sub-300ms stream processing and 92% faster performance than Apache Flink, helping financial institutions block fraud before transaction settlement to prevent $33B annual losses.

Intuit scaled GenAI development across 8,000+ developers with 3,500+ production experiments using the GenOS platform and 'fixed, flexible, free' framework, featuring LLM-as-a-judge evaluation and Agent-friendly API design.

By building a custom MCP server integrated with Amazon Bedrock AgentCore Memory, Kiro CLI achieves persistent conversational memory across sessions, solving the context forgetting issue in Agentic IDEs.

Sovereign AI demands full control over data flow, models, infrastructure, and operations; otherwise GDPR and EU AI Act compliance fails.

Moore Threads released MT Lambda, China's first full-stack domestic embodied AI simulation platform, achieving the first Sim-to-Real verification entirely on domestic hardware and bridging the full link from training to deployment.

Agora-1 lets four humans and AIs battle in the same world-model-generated FPS in real time without any game engine, proving that decoupled simulation-plus-rendering networks can keep multi-client consistency.

Google AI Edge Gallery introduces three major capabilities: MCP protocol support for cross-data-source tool calling, local notification scheduling for proactive interactions, and persistent chat history, shifting mobile Agent development from reactive to automated and continuous experiences.

Google unveiled Gemini Omni video generation model and Gemini 3.5 Flash coding model at I/O, while upgrading AI search to Agent mode and launching Gemini Spark personal AI assistant, marking Google's strategic shift from traditional search engine to AI-powered information Agent platform.

The US defense industry faces a generational opportunity as traditional supply chains remain fragile and unprofitable; autonomy is the key to bending the cost curve—first-principles software-driven design can reduce destroyer construction from 7-9 million labor hours to just 50,000.

Google Cloud's Data Agent Kit is an open-source toolkit integrating BigQuery and other platforms into IDEs or CLIs via MCP, enabling intent-driven data engineering with pre-built skills to solve context window limits and fragmentation.

Google DeepMind's Co-Scientist is a multi-agent AI system designed for scientists that mimics real research team roles to review literature, generate hypotheses, and evaluate ideas, aiming to solve information overload and accelerate scientific discovery.

Amazon SageMaker Feature Store released three new capabilities: native AWS Lake Formation integration for fine-grained access control, new Apache Iceberg table properties to manage metadata lifecycle and reduce storage costs, and a lighter, faster development experience via SageMaker Python SDK v3.

Hugging Face releases the Ettin Reranker Family, six CrossEncoder models ranging from 17M to 1B parameters built on ModernBERT encoders, using distillation training to achieve state-of-the-art performance on MTEB retrieval benchmarks for RAG systems.

Google has launched Gemini Spark, a personal AI agent running 24/7 in the cloud on Gemini 3.5 Flash and Antigravity, autonomously executing cross-app tasks across Gmail, Calendar, and Drive while requiring user confirmation for critical actions, redefining digital workflow automation.

Google's Gemini Omni is the first natively multimodal model for video understanding and generation, enabling arbitrary combinations of image, text, video, and audio inputs with conversational editing and physics-aware reasoning, significantly outperforming prior models like Veo.

The challenge in enterprise AI lies in data infrastructure, not the model. By unifying episodic, semantic, procedural, and workflow memories, Elasticsearch replaces independent systems like Redis and Pinecone, serving as a unified backend engine that significantly reduces operational complexity.

The 2023-2024 El Niño will release 11 million TWh of excess oceanic energy, giving a 99 % chance of the hottest year on record, slashing harvests in Brazil and India and igniting worldwide food-price spikes and South-Asian unrest.

Google announced a transition from assistive AI to autonomous agents at I/O 2026, highlighting the Gemini 3.5 model series, upgraded Antigravity 2.0 agent-first platform, and new tools including Android CLI, Android Bench, and WebMCP to help developers build high-quality applications.

Google AI Edge introduces LiteRT-LM, an optimized inference engine for deploying Gemma 4 models on edge devices, supporting Android, iOS, and web platforms with GPU inference reaching 76 tokens/sec and Multi-Token Prediction delivering up to 2.2x speedup.

Augment Code's benchmark shows that its AI coding assistant Auggie achieves a slightly higher pass rate (67.4% vs 66.3%) than Claude Code using Opus 4.7 while costing approximately 33% less, primarily due to token efficiency from precise retrieval through its Context Engine semantic indexing technology.

Unity Catalog prevents unauthorized AI agent operations through fine-grained access control, audit logging, and AI-specific governance frameworks, ensuring secure and controllable enterprise AI applications.

Databricks announces the General Availability of SQL Alerts, transforming manual data monitoring into automated workflows. The feature lets teams define metrics in SQL, evaluate on schedules, and notify via multiple channels when conditions breach thresholds, with over 4,000 customers already using it in production.

OpenAI announces new content provenance initiatives including C2PA compliance, Google SynthID watermarking partnership, and a public verification tool, building a multi-layered tamper-resistant content traceability system to enhance trustworthiness and transparency of AI-generated content.

Anthropic released Claude Routines on April 14, 2026, the first tool enabling non-developers to create autonomous AI agents in under 5 minutes without servers, code, or continuous operation. Unlike deterministic automation tools like N8N, Claude Routines dynamically adjusts reasoning paths based on input changes, supporting both local and remote execution modes.

TanStack disclosed a sophisticated npm supply chain attack that compromised 42 packages, with attackers injecting malicious code by hijacking maintainer accounts and exploiting npm publishing process vulnerabilities—a major security incident targeting the JavaScript ecosystem in 2026.

Agoda builds a multimodal content system using AI technology to semantically connect user-uploaded images with hotel reviews, enabling intelligent upgrade of travel discovery experience and helping users more intuitively evaluate accommodation options.

MTP is a new speculative decoding feature built into llama.cpp that can approximately double token generation speed for most use cases, achieving ~30 tok/sec with the Dense 27B model and ~100 tok/sec with the MoE model.

Video search no longer relies on transcripts or metadata; it now directly embeds video clips via multimodal models for retrieval.

Google I/O '26 introduces Antigravity 2.0 and the Managed Agents API, establishing a four-rung agent development spectrum from low-code to full code control, unified by the A2A protocol.

The public sector is entering the agentic era, where AI agents compress drug reviews from months to hours, migrate billions of emails in 22 days, and handle mega-events with millions of visitors.

Google DeepMind's WeatherNext AI model excelled in predicting 2025's Hurricane Melissa, accurately forecasting its rapid intensification to Category 5 and landfall path 3 days in advance, securing critical evacuation time for Jamaica and proving the practical value of AI in extreme weather warning.

Researchers in Uganda used AlphaFold to narrow down breast cancer vaccine targets from 15,000 to 15, significantly lowering research barriers and advancing local precision medicine.

AI tools are revolutionizing antibiotic development by accelerating structure elucidation and discovering non-intuitive patterns to combat the silent pandemic of antimicrobial resistance.

Building AI Agents should follow four maturity levels: validate with frameworks, customize with state machines, optimize UX with Kanban, and deploy to the cloud. Avoid hype and evolve from simple to complex based on needs.

AWS proposes three architectural patterns for voice agents—direct tool calling, agent-as-tool delegation, and session segmentation—built on Amazon Nova Sonic, Bedrock AgentCore, and Strands BidiAgent, achieving sub-500ms end-to-end latency while solving multi-agent coordination and real-time audio streaming challenges.

At the AMD AI Developer Conference in Shanghai, CEO Lisa Su stated that AI competition is shifting from model capabilities to systems engineering and full-stack optimization. Developers need a deployable, optimizable, and continuously evolving engineering system. AMD, centered on its ROCm open-source platform, provides full-stack computing power from cloud to edge, while continuously strengthening its developer ecosystem in China.

KPMG establishes a global strategic alliance with Anthropic to deeply integrate Claude into its core Digital Gateway platform, granting access to over 276,000 employees to enhance efficiency in tax, legal, and cybersecurity sectors.