AI 今日新闻 · 2026-05-23

Octopus Energy reduced margin data engineering costs by 50x to handle 48x data growth from UK’s MHHS policy, using a three-stream architecture and Delta Lake CDF for incremental processing—cutting daily rows from 25B to 300M.

Virgin Atlantic used Codex to launch its new mobile app before the high-risk Christmas rush, achieving zero P1 defects and near-full test coverage, while accelerating legacy refactoring by up to 80% and enabling non-engineers to build data apps in hours.

NVIDIA introduces Nemotron-Labs Diffusion, a diffusion-based language model family supporting AR, diffusion, and self-speculation modes, enabling parallel multi-token generation and iterative refinement for higher throughput and flexibility.

Anthropic's Project Glasswing has discovered over 10,000 high/critical-severity vulnerabilities in one month using Claude Mythos Preview, with Mozilla finding 271 in Firefox 150—10× more than with Opus 4.6—marking AI-driven vulnerability discovery entering large-scale practical deployment.

The TeamPCP group has launched over 20 waves of supply chain attacks in months, compromising more than 500 distinct open-source tools—over 1,000 versions—and exposing at least 3,800 internal GitHub repositories.

An internal OpenAI model disproved the famous Erdős unit distance conjecture, demonstrating AI's breakthrough capability in mathematical research.

Enterprise RAG systems should focus on document understanding and business logic rather than stacking models and frameworks. Simple Python scripts often outperform complex production systems.

Google launches ADK for Kotlin and Android 0.1.0, enabling privacy-preserving AI agents on mobile devices with cloud-edge collaboration.

AI Desk Card is an intelligent device that uses an e-paper screen and AI agent to manage desktop information, replacing traditional sticky notes for better productivity.

This article systematically explains how to build a multi-agent research assistant using OpenAI Agents SDK, GPT-5.4 mini, and Olostep Web API, enabling automated web search, evidence evaluation, and structured report generation.

Databricks introduces a production-ready AI agent observability solution using Unity Catalog and OpenTelemetry, storing traces as Delta tables in the Lakehouse to enable long-term retention, SQL analytics, PII governance, and MLflow-based evaluation loops.

OpenAI has been recognized as a Leader in the Gartner® Magic Quadrant™ for Enterprise AI Coding Agents, driven by Codex’s support for 4M+ weekly users, GPT-5.5 integration, enhanced tool use, and enterprise-grade governance capabilities.

AI coding tools have evolved beyond autocomplete, but complex feature development requires structured workflows—not just better prompts. A five-layer system (context, knowledge, agents, workflows, delivery) enables scalable, reliable AI-assisted engineering.

This article systematically introduces how to build an AI-powered medical image de-identification pipeline that removes PHI from both DICOM metadata and burned-in pixel text via OCR + NER + DICOM processing, ensuring HIPAA compliance.

AtomBite.AI, founded by former Meituan Waimai tech lead Dr. Wang Dong, focuses on embodied AI for commercial kitchens using World Action Models (WAM), securing seven-figure seed funding and multiple pilot deployments with top-tier food service companies.

Product experimentation for collaborative AI features faces user interdependence issues, where traditional user-level A/B testing fails due to shared artifact propagation, workflow interference, and network adoption, requiring cluster randomization to solve the collaborator contamination trap.

Specialized models outperform large general models in enterprise applications, with DharmaOCR's 3-billion-parameter specialized model defeating all commercial frontier APIs in structured OCR tasks at one-fiftieth the cost. Enterprise AI procurement strategies should shift from scale-oriented to specialization-oriented approaches.

Li Fei-Fei's team releases ESI-Bench evaluation benchmark, transforming observers into actors for the first time, closing the perception-action loop, and discovering three major issues in current AI spatial intelligence: action blindness, 3D reconstruction traps, and metacognitive deficits.

Google launches full-stack AI capabilities for Gemini for Home, empowering service providers and hardware partners to build intelligent home solutions.

The article dives deep into how Linear achieves high performance through IndexedDB, initial load optimization, and optimistic updates, offering significant engineering insights.

Gemma 4 enables true agentic behavior through local sandboxed tools like filesystem exploration and restricted Python execution.

Google I/O emphasized that the path of AI in science is shifting from specialized tools to general agent systems, signaling a fundamental change in research paradigms.

Even assuming AGI requires a new paradigm, applying Lindy's Law suggests it may emerge within 3 to 5 years, so current AI development risks shouldn't be underestimated.

Google releases Gemini Omni, an advanced AI video model supporting multimodal inputs and physics understanding, filling the gap left by Sora's shutdown.

This video explains the foundational elements of AI chip design, from logic gates to matrix multiply-accumulate operations, revealing core hardware mechanisms behind AI computation.

This article proposes a new sequence-level load balancing method called Moving Quantile Balancing (MQB), which achieves fine-grained balance in MoE models without relying on auxiliary losses.

AI amplifies the divide between law and logic; enterprises need frameworks to translate legal intent into machine-executable controls.

Enterprise AI systems must combine deterministic analytics with LLM reasoning; relying solely on large models leads to erroneous results.

The article highlights how memory costs heavily impact vector search operations and introduces six techniques provided by Milvus to reduce memory pressure.

Memory shortages are driving price increases in consumer electronics, especially impacting budget smartphones significantly.

Agentic programming is a paradigm where AI models act as autonomous decision engines inside software systems—executing workflows rather than just responses—yet only 11% of enterprises run agents in production, mainly due to engineering and architectural gaps, not lack of demand.

Japanese firms commonly pursue extreme diversification—not randomly, but via deep material-science roots—as illustrated by Toto, a toilet maker that unexpectedly became a top supplier of e-chucks for AI memory chips.

AI development is constrained by physical infrastructure, giving ordinary people veto power over AI projects through data center approvals, creating new leverage against tech giants.

The core bottleneck in quantum machine learning lies in efficiently encoding classical data into quantum states, with no universally efficient loading methods currently available.

This article explains how to build a context-aware semantic search engine in Python using LLM embeddings combined with metadata filtering.

ChatGPT Workspace Agents provide enterprise-level control features, allowing administrators to manage agent builder permissions through role-based access controls and set application access restrictions and human confirmation flows to ensure secure operations.

Cloudflare extends its cloud access security broker (CASB) to support Anthropic's Claude Compliance API, allowing security teams to monitor Claude usage directly in the Cloudflare dashboard without endpoint agents. This addresses the issue of AI application security governance lagging behind adoption, providing full lifecycle security coverage from API calls to data processing.

Discord rebuilds its database operations around automation to manage large-scale ScyllaDB clusters, achieving zero-downtime upgrades, automatic failure recovery, and elastic scaling through self-developed operation platforms to support real-time communication services for millions of users.

Rider 2026.2 EAP 3 introduces AI agent test coverage skills that reduce token consumption by 50% using built-in coverage data, adds code change preview functionality, game development templates, and improved NuGet management experience.

JetBrains Rider 2026.2 introduces AI agent skills system that helps AI intelligently locate test positions through code coverage data, reducing AI costs by half and solving the efficiency problem of AI agents blindly searching project files.

Movix has developed a specialized AI solution for dental appliance manufacturing, using custom AI models and the Gemini Enterprise Agent Platform to automate quality control, digitize traditional manual processes, and significantly improve production efficiency and product quality.

AI-driven cyberattacks are reshaping the security landscape, requiring traditional SOCs to transform into agentic SOCs to counter machine-speed threats. Attackers have率先 adopted AI technologies, reducing average breakout time to 29 minutes, forcing defenders to use AI against AI to maintain balance.

Google announced several AI updates at I/O 2026 including the faster and cheaper Gemini 3.5 Flash and the powerful multimodal model Gemini Omni, sparking community debate.

Modular monolith architecture improves code maintainability through clear boundaries and layered design, suitable for enterprise applications requiring long-term evolution.

Models.dev is a community-driven open-source database that aggregates AI model specs, pricing, and features, offering a standardized API and TOML-based data format for easy integration and updates.

Leading AI companies are shifting from pure model development to end-to-end agent systems, with OpenAI, AI21, and DeepSeek all forming Agent/Harness teams—marking a paradigm shift from ‘models as product’ to ‘systems as product’.

The World Bank Group built a unified data and AI platform on Databricks to integrate structured and unstructured knowledge for the first time, enabling intelligent Q&A and outcome-driven decision-making across global development projects.

Firebase Cloud Functions now offers experimental Dart support, enabling Flutter developers to write both frontend and backend in the same language, share data models and validation logic via a common Dart package, and reduce full-stack development friction significantly.

Tie 100 ends of 50 strings randomly in pairs until no ends remain; the expected number of loops is the 50th odd harmonic number ≈ 4.5.

AI Agent era demands a new data stack: legacy ETL-based architectures cannot handle agents’ real-time, high-frequency, multi-source access and security requirements.

AI agents are non-deterministic and interactive, making traditional testing methods inadequate; simulation sandboxes are essential to robustly validate behavior before production.

AI evals are fundamentally broken—over-reliance on objective metrics misleads—but they remain critical when built, interpreted, and embedded properly in agent workflows.

Cogent builds a three-agent autonomous cyber defense system that responds at machine speed to counter the compressed vulnerability exploitation window—now down to minutes—processing billions of events daily with hot/cold context separation and LLM judge validation.

DeepSeek permanently applies a 75% discount to V4 Pro pricing and reduces cache-hit input prices to 1/10 of original for all models, bringing V4 Pro input cache-hit cost to $0.003625/1M tokens.

AI is reshaping programming, scientific research, and sports competitions, pushing new frontiers in automation and human enhancement.

Three AI infrastructure companies Exa, Modal, and TurboPuffer have achieved unicorn status with valuations reaching $2.2B, $4.7B, and $100M ARR respectively, while RAEv2 achieves 10x faster convergence and NVIDIA's Gated DeltaNet-2 surpasses KDA and Mamba-3 at 1.3B parameters.

Google I/O 2026 reveals Gemini's integration into every product as AI agents, with massive scaling from 9.7 trillion to 3.2 quadrillion monthly tokens, new TPU chips split for training/inference, and Gemini Omni as the headline multimodal model capable of understanding reality through text, video, and sound inputs.

Eric Ries' new book Incorruptible reveals how great companies prevent mediocrity and corruption, emphasizing that successful companies need to actively protect their core values rather than relying on success itself to ward off threats.

Cisco leveraged Codex AI to reduce AI defense feature development cycle from several quarters to weeks, with engineers using Codex to write 100% of new feature code and build the open-source tool Defense Claw. AI-assisted programming significantly enhanced engineer ambition and productivity, changing project evaluation frameworks.

Anker noise-canceling headphones achieve flagship-level audio quality, noise cancellation and call performance in the 1000-1500 yuan price range through self-developed Thus™ compute-storage integrated AI chip, integrating AI recording and translation functions, marking Anker's transformation from product integrator to core technology controller.