AI 今日新闻 · 2026-05-31

AI inference theft is extremely costly—single calls can hit $2—attackers use proxy adapters to steal at scale; Vercel deploys BotID for deep analysis, and developers can integrate it in minutes.

Databricks Lakebase enables Martin Fowler’s long-aspirational Practice #4—“every developer gets their own database instance”—via copy-on-write branching, making evolutionary database development operational at production scale for the first time.

OpenAI proposes a universal framework for trustworthy third-party evaluations, emphasizing that reports must explicitly state the claim being tested, provide validity evidence, distinguish three claim types (capability elicitation, safeguard performance, comparison), and recognize that the 'harness' critically shapes evaluation outcomes for long-horizon tasks.

AlloyDB Hot Standby reduces failover time from minutes to ~15 seconds and eliminates performance degradation from cold cache warm-up—all at zero additional cost; the standby node continuously applies WAL logs, enabling near-real-time synchronization with the primary.

A2UI is an open protocol enabling AI agents to safely return structured UI components (e.g., date pickers, maps) instead of plain text; integrated with Gemini Enterprise, it renders rich, interactive interfaces natively in chat surfaces—and supports cross-framework (Lit/Flutter/Angular) and transport-agnostic (A2A/SSE/WebSocket) deployment.

BYD launches China’s first automotive-grade 4nm AI driving chip Xuanji A3—three chips combined exceed 2100 TOPS, 20% lower power per TOPS, 100% higher compute utilization, now mass-produced.

Gamma-World systematically solves architectural gaps in multi-agent world modeling via simplex agent encoding and sparse hub attention, achieving >40% average FVD reduction, zero-shot generalization from 2 to 4 agents, and 24 FPS real-time rollout.

AWS proposes a full-stack observability solution for SageMaker LLM inference, collecting infrastructure metrics (GPU utilization, latency) and custom quality metrics (response accuracy, compliance) via CloudWatch, visualized in Managed Grafana—enabling dual-dimension monitoring to address cases where systems appear healthy but produce poor outputs, or deliver high-quality responses inefficiently.

Endeavour Energy rebuilt its real-time outage map using Next.js and Vercel, achieving sub-1s loads, 5-min data sync, and 38% faster deployments — solving infrastructure bottlenecks during storm surges.

Starting Jan 1, 2026, 700+ U.S. hospitals must manage total cost and quality across 5 high-cost surgical episodes over 30 days; legacy analytics cannot enable proactive intervention—success hinges on unified lakehouse platforms, embedded AI workflows, and scalable architecture, or 66% will lose revenue due to data latency.

Rust IDEs (e.g., rust-analyzer and RustRover) must reimplement the compiler frontend—not reuse rustc—to achieve low-latency interactivity; their core challenge lies in handling incomplete code, incremental parsing, and on-demand semantic analysis.

Connected Sheets enables direct integration between Google Sheets and BigQuery, allowing business users to analyze petabyte-scale data in real time without SQL—eliminating data silos and security risks from CSV exports; proven use cases include billion-row pivot tables, auto-refreshing dashboards, and hybrid modeling.

Zed trained Zeta2 via production edit data distillation: using frontier models to generate candidate edits, filtering low-quality outputs with static evaluation and a 'repair' mechanism, yielding ~100K high-quality training examples; the entire pipeline is JSONL-based for fast experimentation.

Particle physicist Don Lincoln states: the near-total absence of antimatter in the universe remains unsolved; dark energy drives cosmic acceleration but its nature is unknown; a Theory of Everything may exist yet remains far from experimental verification; while the Standard Model is complete, it fails to explain gravity or dark matter, and scientific progress hinges on testable predictions—not mathematical elegance.

DoorDash built a 'simulation and evaluation flywheel' system that uses offline realistic multi-turn conversation simulation and automated grading to reduce LLM chatbot hallucination fixes from weeks to hours, dramatically improving iteration speed and deployment confidence.

The author open-sourced the review-forge toolchain, which controls AI-generated code drift via multi-model cross-review, consensus synthesis, human-driven prioritization, and AI-based fix-verify loops.

This beginner-friendly guide walks through using torch.profiler to analyze a matrix multiplication + addition operation, revealing CPU-GPU coordination patterns and how torch.compile fuses operations to reduce kernel launch overhead.

Salesforce's engineering team evolved from relying on Copilot to building an Agentic engineering system, using three key levers—tool convergence, rule-as-code, and autonomy—to delegate SDLC execution to Agents, achieving a 79% increase in PRs, 151% higher effective output, and completing a 231-person-day API migration in just 13 days.

The layoff wave is a corporate excuse to avoid strategic transformation; AI efficiency narratives mask the lack of organizational and business restructuring—what’s needed is creating AI-native businesses, not cutting headcount.

NVIDIA VSS and NemoClaw enable engineers to deploy a video search AI agent in 5 minutes without writing integration code, achieving fusion search via natural language queries for accurate results.

AI Forward-Deployed Engineers (FDEs) are resurging due to demand for custom agentic workflows, but AI Engineer roles will vastly outnumber FDEs; enterprises prefer building in-house AI teams to preserve vendor optionality, with current high demand focused on generalist AI Engineers skilled in LLM prompts, agentic frameworks, evals, and AI coding agents.

Nsight Copilot runs offline on DGX Spark using 128GB VRAM to deploy GPT OSS 12B NIM + CUDA RAG pipeline, delivering privacy-preserving, cloud-cost-free AI coding assistance for CUDA developers.

Claude Opus 4.8 is currently the strongest coding model available, but its API is expensive ($5/million input tokens, $25/million output tokens); Verdant offers a 7-day free trial with no credit card required, supporting multi-Agent parallel development, isolated Git workspaces, and Plan-First workflows to significantly improve coding reliability and engineering control.

David Reich proposes that Neanderthals may not be an evolutionary side branch, but rather a product of modern human cultural expansion: a population inventing the Middle Stone Age spread into Europe and Africa, mixed with local archaic humans, retained modern culture despite ~95% genetic replacement, sharing Y chromosomes, mtDNA, and toolkits—making them our ‘cultural cousins’.

Terence Tao states that AI is significantly reducing cognitive friction in mathematical research, enabling more efficient experimentation, collaboration, and literature search; he emphasizes AI tools are now 'ready for prime time' and advocates sharing exploration paths to enhance collective knowledge accumulation.

Senior engineers struggle with AI Agent development because the paradigm shifted from deterministic programming to iterative prompt-feedback loops; text is now the new state representation, and engineers must transition from 'traffic controllers' to 'dispatchers'.

Public-sector CISOs should build AI-ready security programs across five core domains in phases: deliver AI-driven board reporting and vendor optimization in 90 days, SOC automation and policy generation in 6 months, and proactive hunting & architecture integration in 6–12 months—using a hybrid strategy of custom workflows (e.g., Gemini Gems), commercial AI procurement (e.g., Gemini for Government), and integration with existing stacks, all on FedRAMP High/DOD IL5-accredited platforms.

Third Point founder Dan Loeb argues AI is reshaping investment logic: investors must grasp the AI Stack (power → chips → models → applications); credit expertise is critical in volatile markets; his $25B multi-asset platform coordinates equities, credit, insurance, and private markets via strategies like Fulcrum Security; activist investing succeeds through governance reform, not confrontation.

Nick Nisi found that by trimming AI agent skills from 95% down to just 5 core roles (implementer, verifier, reviewer, closer, retro), he achieved higher-quality outputs; the key was replacing Claude-native skills with a TypeScript state machine to solve context loss.

Yi Tay, a Google DeepMind research scientist, was the core contributor to Gemini Deep Think achieving IMO gold in 2025; he co-founded Reka AI and trained a GPT-4-level multimodal model, with key technical contributions including UL2, DSI, and PaLM-2—and holds an Associate Diploma in classical piano from Trinity College London.

OpenAI’s official free ChatGPT add-in is now integrated into PowerPoint desktop apps, enabling natural-language-driven creation, updating, understanding, and polishing of presentations—all fully editable as native text boxes, with no cost and better performance than Microsoft Copilot.

Production-grade Agent Harness cannot be solved by framework selection alone; it must systematically handle 15 core responsibilities (e.g., policy, approval, budget, trace), each implemented as an installable, versioned, language-agnostic worker.

LangChain introduces Managed Deep Agents, a customizable agent harness architecture supporting complex real-world tasks via execution environment, context management, delegation, and human-in-the-loop capabilities.

Vercel Sandbox now supports running Docker containers within an isolated environment, enabling developers to build images, install system packages, and deploy apps without affecting the host machine.

Mole CLI v1.40.0 is released with features to reclaim multi-GB git worktrees, clear caches for Chrome DevTools/Spacedrive/QQ Music, fix Dock/wallpaper issues, diagnose battery health, and speed up orphan scans 15x.

Cloud code agents like Claude Code may delete databases without explicit instruction if lacking context-aware controls; existing API security tools cannot intercept such high-risk actions due to inability to interpret their decision logic.

brew-browser is a lightweight native macOS GUI app built with Tauri 2 and Svelte 5, adding dashboard, package search/install/uninstall, service management, and Brewfile snapshot features to Homebrew—greatly improving CLI usability.

Herdr is a lightweight, Rust-based terminal-native multi-Agent management tool supporting tmux-level persistence, Agent state awareness, and parallel execution—no GUI required, config-inherited, ideal for resource-constrained environments.

Claude Opus 4.8 is Anthropic’s rapid revision of the controversial 4.7 model, prioritizing improved ambiguity handling to restore the user-friendly ‘vibes’ of 4.6; though it outperforms GPT-4.5 on official benchmarks, real-world engineering benchmark DeepSWE shows GPT-4.5 currently leads—and 4.8 hasn’t been tested yet.

Google delayed Gemini 3.5 Pro to June and pivoted to promoting Flash; real-world tests show Flash excels at occluded object recognition and native video understanding (timestamped insights + Python chart generation), effectively replacing Pro-tier capabilities.

This project builds a terminal TUI dashboard using Bun + OpenTUI + SolidJS to unify Kanban task boards, calendar scheduling, and Claude Code agent sessions—though it lacks technical depth or architecture details.

Gemini’s free tier is currently the most generous among major AI providers—especially in Google Search AI mode, which rarely throttles even on free plans; its $100/month plan bundles YouTube Premium, 20TB storage, and 10,000 Flow Credits, offering exceptional value.

Xiaohongshu’s acquisition of World Cup rights is not merely content expansion, but a strategic pivot from a ‘recommendation community’ to a broad-interest platform—aimed at addressing male-user deficit and long-video capability gaps, yet facing community-tone conflicts and uncertain ROI.

Vercel is switching function invocation billing from package-based to per-unit pricing for Pro and new Enterprise customers, with existing customers retaining current rates until end of billing cycle.

Apple is reportedly using Gemini to train on-device AI and partially relying on Google Cloud; Huang Renxun says he’ll ‘explode’ if engineers spend less than $250K/year on tokens (vs. $500K salary); MIIT confirms EV batteries have entered mass-scale retirement; enterprise AI bills can hit $500M/month.

Google announced Gemini AI glasses—audio-only version launching this fall, display version still in prototype—with Warby Parker, Gentle Monster, and Samsung; all AI functions require phone pairing and lack on-device inference.

Gary Marcus highlights severe intellectual homogeneity in AI commentary and research; though improved recently, the ecosystem remains unhealthy—sparking peer reflection on communication strategies.

The current 'find out stage' of AI is fundamentally about supply chain security and password protection—not algorithmic breakthroughs; Dataiku stresses structured frameworks and reusable data products, while 1Password highlights identity standard gaps for ephemeral agent swarms.

Google Cloud unilaterally suspended Railway’s production account, triggering an 8-hour platform-wide outage—highlighting critical risks of single-cloud dependency and absence of failover mechanisms.

Meta did not disclose technical details; the page is merely an InfoQ navigation/ad template with no substantive content—only a headline claiming PB-scale ingestion rebuild.

Arm open-sourced the AI security framework Metis, claiming superior performance over traditional SAST—but the article lacks technical details, benchmarks, or code links, making it a low-density press release.

Luis Garicano argues that if OpenAI and Anthropic face a ~4-month model iteration lag (per EpochAI), they may achieve zero profit despite revenue growth; Marc Andreessen retweeted with 'Fight! 🦾' signaling support.

OpenAI Developers proposed a voice-first mobile OS where users speak, agents respond, and execute cross-app actions—but the post contains only one sentence + two links, zero technical details.

Guangfan Tech has partnered with Tencent Mobility Services to integrate its AI-powered wearable device into the latter’s mobility platform, with features expected to launch in early June; the device topped JD.com’s AI headset bestseller list for 8 consecutive days and sold out its first batch, triggering a new pre-sale round.

ChatGPT Images 2.0 can generate full-page Song Dynasty aesthetic-style PPT visuals, and the GitHub tool KK-C automates packaging them into PPTX, PDF, and web-based PPT formats—yet the post only includes a tweet screenshot and link, lacking technical details.

DeepSeek-V4 Pro is praised for cost-effectiveness in small tasks like code review and writing, replacing expensive Qwen-Max; current primary model ranking: GPT-5.5 > Claude 4.7 > DeepSeek-V4 Pro.

NVIDIA Dynamo collaborates with Gcore and Orange Business for scalable distributed AI inference, but the article is only a YouTube video link with no substantive content—no architecture, metrics, or implementation details provided.

Gary Marcus argues Anthropic has a higher probability of long-term profitability than OpenAI, yet both remain low; a commenter adds its path via Claude, Cowork, and enterprise software is razor-thin.

Corporate AI infrastructure often lacks clear ownership, leading to shadow IT, 2 a.m. crises, and AI gaslighting — author will host an unfiltered panel at Cisco Live to address these issues.

Open-source LLMs face three structural barriers: unsaturated pre-training, $2–4B training costs per model, and blocked distillation due to closed APIs—Andreessen signals solidarity with 'Fight! 🦾'.