本周 AI 回顾 — 2026年6月22日 – 6月28日

NVIDIA NeMo AutoModel significantly accelerates fine-tuning Mixture-of-Experts (MoE) models by building on HuggingFace Transformers v5. It integrates Expert Parallelism, DeepEP fused all-to-all dispatch, and TransformerEngine kernels, leveraging v5's dynamic weight loading. This results in 3.4-3.7x higher training throughput and 29-32% less GPU memory compared to native Transformers v5, using the same API. NeMo AutoModel enables efficient scaling of MoE models, even for frontier-scale models where v5 runs out of memory.

OpenAI and Broadcom have collaborated to launch Jalapeño, a new custom AI chip. This chip is specifically designed to optimize large language model (LLM) inference. The goal of Jalapeño is to enhance the performance, efficiency, and scalability of AI systems, addressing key areas for advancement in artificial intelligence.

PP-OCRv6, PaddleOCR's latest universal OCR model, offers three tiers (1.5M to 34.5M parameters) for diverse deployment. The small and medium tiers support 50 languages. It achieves 86.2% detection Hmean and 83.2% recognition accuracy with the medium tier, improving upon PP-OCRv5. Key advancements include PPLCNetV4 backbone, RepLKFPN for detection, and EncoderWithLightSVTR for recognition, ensuring accurate, structured text output across various real-world scenarios.

OpenAI is launching "Health in ChatGPT" for U.S. users, allowing secure connection of Apple Health and medical records to help users understand their health information. This feature enables personalized conversations and tracking of changes, with users controlling data access. The system's performance is evaluated using HealthBench Professional, where GPT-5.6 models surpassed GPT-5.5. Users can review and update connected information, but should verify important details with original sources as data may not always be current.

BuzzRadr reports on GLM-5.2, a new flagship model for long-horizon tasks, featuring a solid 1M-token context. It introduces advanced coding with flexible effort levels and an improved architecture, including IndexShare, reducing FLOPs by 2.9x. GLM-5.2 is open-source under an MIT license and demonstrates strong performance on coding benchmarks, often outperforming competitors and ranking as the highest open-source model.

New research in "Nature" highlights AMIE, an AI system, for its potential in long-term health condition management. Utilizing Gemini models, AMIE features an empathetic dialogue agent and a reasoning agent that cross-references clinical knowledge. In a blinded study with patient actors, AMIE matched physicians in management reasoning and surpassed them in plan preciseness and guideline alignment, suggesting its future role in supporting medical care and freeing up physician time. Further studies are exploring its use in clinical settings.

OpenAI research reveals AI agents are revolutionizing work by facilitating longer, more intricate tasks. This advancement significantly boosts productivity across various job functions, demonstrating a transformative impact on the modern workplace.

The FFASR Leaderboard, a collaboration between Treble Technologies and Hugging Face, is the first open, community-driven benchmark for evaluating ASR models in realistic far-field acoustic conditions. It addresses the significant gap between clean-speech benchmarks and real-world performance, where reverberation and noise degrade accuracy. The leaderboard uses hybrid wave-based simulation across 14 diverse rooms, validated against real measurements, to assess models' robustness. It tracks Word Error Rate (WER) and RTFx, encouraging development of models suited for complex environments like AI voice agents and in-car assistants.

Immunologist Derya Unutmaz utilized GPT-5 Pro to unravel a three-year-old mystery concerning T cell behavior. This significant advancement holds promise for future research in both cancer and autoimmune diseases. The AI's contribution provided crucial insights, potentially accelerating scientific understanding and therapeutic development in these critical areas of medicine.

Google Search tools are revolutionizing thrift and vintage shopping, with "vintage" and "how to thrift" trending. AI Mode helps plan shopping trips, even suggesting gluten-free brunch spots. Google Lens uncovers hidden gems by identifying items and their value. Circle to Search finds similar items online from images. Virtual Try-On lets users digitally preview vintage clothes. Additionally, Lens can help users assess the resale value of their own items, promoting a circular economy.

Hugging Face now ships weekly updates for `huggingface_hub`, its core Python client, using an automated workflow. This system, built with open-source tools like GitHub Actions and an open-weights AI model (GLM-5.2), automates mechanical tasks like version bumping and publishing. A human remains in the loop to review and edit AI-drafted release notes and announcements, ensuring accuracy and quality. This approach aims for reusability, allowing other maintainers to adapt the workflow.

Google I/O 2026 leveraged AI tools like Gemini and Nano Banana to create various elements, from a "TPU Training Day" film blending human artistry with experimental technology, to the visual brand identity. AI also powered immersive pre-show experiences, including "Jellectronica," which translated jellyfish movements into music, and "Infinite Scaler," a video game where players generated levels using AI. This showcased how AI can enhance creativity and efficiency in event production.

OpenAI has launched Patch the Planet, a new Daybreak initiative. This program aims to assist open-source maintainers in identifying, validating, and resolving security vulnerabilities. It leverages a combination of artificial intelligence and expert review to support these efforts.

Google's I/O 2026 announcements are highlighted in a new quiz, "vibe coded" in Google AI Studio. This quiz, built by an editor with no coding experience using Gemini and Google AI Studio, demonstrates how the Antigravity coding agent allows users to create with the latest Gemini models. The quiz showcases new applications for Google AI Studio, encouraging developers and builders to explore its capabilities.

Google I/O 2026 introduced Gemini Omni and Gemini 3.5. Gemini Omni creates and edits videos from various inputs using natural language, allowing users to reimagine scenes and refine content. Gemini 3.5 Flash, a new model family, excels at agentic tasks, coding, and powers interactive web UIs. It's now the default for the Gemini app and AI Mode in Search, enabling features like information agents and custom generative UIs, with Gemini Spark acting as a personal AI agent integrated with Workspace tools.

Transformers.js allows Web developers to integrate AI models. A challenge arises with caching: even if multiple applications use the same AI models or Wasm runtime files, browsers re-download and re-cache them if they originate from different top-level sites. This is due to cache isolation, a security measure preventing timing attacks by partitioning caches based on the top-level site and current-frame site, leading to redundant downloads and storage.

Samsung Electronics is rolling out ChatGPT Enterprise and Codex to its global workforce. This deployment represents one of OpenAI's most significant enterprise AI implementations to date, providing Samsung employees with access to these advanced AI tools.

Google Finance has released new updates, including an Android app, to help users track investments and stay informed. Users can now consolidate portfolios, gain insights into asset allocation, and utilize a research tool. The platform also offers customized market intel briefings based on user preferences. The new Android app provides real-time data, news, and an AI research tool, with an iOS app planned for later this year.

Users can launch a private, OpenAI-compatible LLM endpoint on Hugging Face infrastructure with a single command. This allows for quick setup of models for testing, evaluations, or batch generation, with billing per-second for hardware usage. The endpoint is gated and requires an HF token for access, ensuring privacy. Users can query the server from various platforms and scale to larger models by adjusting hardware flavors and parallelization settings.

Google is investing $1.5 billion to expand its Jackson County, Alabama data center by 2027, funding its own power and infrastructure. This expansion includes a $2 million Energy Impact Fund with TVA and CAANEAL for local energy efficiency. Google is also donating $550,000 for STEM kits for students, building on its existing community efforts like water stewardship, digital skills training for over 130,000 Alabamians, and job creation.

Google is expanding its commitment to Virginia with new community investments. These initiatives aim to create thousands of local jobs and prepare the future workforce by funding electrical apprenticeship training, targeting 2,741 additional apprentices by 2030. Furthermore, Google is launching a $15 million Energy Impact Fund to reduce utility bills for Virginians through home repairs and energy-efficiency upgrades, alongside investing in over 500 megawatts of new energy capacity.

OpenAI is actively involved in establishing shared standards for advanced AI. They contribute to this effort by supporting the development of evaluation frameworks and safety practices. Furthermore, OpenAI promotes global cooperation in the AI field through its involvement with the Appia Foundation, aiming to foster a unified approach to AI development and deployment.

BuzzRadr Trending Topics: A new benchmark evaluates how efficiently open models, specifically coding agents, utilize software libraries like transformers. Traditional benchmarks only assess final answers, but this new approach measures the entire process, including the effort an agent expends. The goal is to optimize library design for agentic use, emphasizing clear APIs, extensive documentation, and agent-specific testing to reduce costs, latency, and token usage. The evaluation uses a harness with the pi coding agent across various models, revisions, and tasks on Hugging Face Jobs.

BuzzRadr reports new community investments in Missouri, coinciding with a new data center in Montgomery County. A Capacity Commitment Framework with Ameren will support over 500 megawatts of additional energy capacity. A $20 million Energy Impact Fund aims to reduce utility bills through home weatherization. The company is also funding workforce programs, including training for construction laborers and apprentices in Montgomery County, to prepare Missourians for skilled roles.

OpenAI has unveiled Deployment Simulation, a new method designed to anticipate AI model behavior prior to release. This technique utilizes real conversation data to simulate how models will perform in actual deployment. The goal is to enhance safety and improve the accuracy of model evaluations before they are made public.

OpenAI has unveiled a preview of GPT-5.6 Sol, their next-generation model. This new iteration promises enhanced capabilities across several key domains, including coding, scientific research, and cybersecurity. A significant feature of GPT-5.6 Sol is its integration with OpenAI's most advanced safety stack, suggesting a strong focus on secure and responsible AI development.

MosaicLeaks reveals deep research agents risk privacy by leaking sensitive information through external queries, even when combining private documents with public tools. A new task with multi-hop questions shows agents frequently leak private data, and training solely for performance worsens this. The proposed Privacy-Aware Deep Research (PA-DR) method improves task success to 58.7% while significantly reducing answer/full-information leakage from 34.0% to 9.9% by training agents to avoid revealing private facts through query logs.

Omio is leveraging OpenAI to revolutionize travel experiences, focusing on conversational AI. This integration is not only accelerating their product development but also fundamentally transforming Omio into an AI-native company. Their strategy centers on using AI to enhance user interaction and streamline their operational processes.

Jason Liu leverages Codex to maintain context and oversee intricate projects, ensuring the continuity of work beyond initial prompts. This approach allows for sustained development and management of long-running tasks.

The Google-funded Futures Lab, a partnership with the University of Waterloo, is developing innovative AI prototypes. Students from various fields participate in eight-week intensive workshops, creating tools to transform

Google Beam is enhancing group meetings with a new experiment. It renders remote participants in true size on HP Dimension's immersive display, positioning them around a virtual table with spatial audio. This aims to boost social connection by 50% and contribution ability by 21%, closing the hybrid "inclusion gap" for users joining from various devices.

AI Mode, launched a year ago, now boasts over a billion monthly active users globally, with queries doubling quarterly. It's changing U.S. search habits, expanding what's searchable. Over 1 in 6 searches use voice or images, and average AI Mode queries are triple the length of traditional searches. Planning and brainstorming queries are growing significantly, indicating users are "getting things done" and "making decisions."

This BuzzRadr summary, "Profiling in PyTorch (Part 2)," details how nn.Linear replaces a hand-written matmul-add pair. It explains that nn.Linear folds bias addition into the matrix multiplication kernel using an epilogue, avoiding separate add operations. The summary also notes that `torch.compile` offers minimal fusion benefits for a single `nn.Linear` operation because the underlying cuBLAS GEMM kernel already handles bias addition efficiently.