AI 日报 — 2026-08-28

Gemini 3.5 Transcribe, launched on August 26, 2026, marks a significant improvement over the previous Chirp 3 transcription model. It offers new capabilities, reduced word error rates, and notably better latency, with a 70% improvement in time to final transcription. The model demonstrates precise multilingual performance on the FLEURS benchmark, achieving a 5.50% WER in streaming mode and 5.04% WER in non-streaming use-cases.

A judge has ruled that the Trump administration's blacklisting of Anthropic was illegal. This decision, reported by nytimes.com, indicates a legal challenge to the previous administration's actions regarding the company. The ruling suggests that the blacklisting did not comply with legal standards, potentially impacting similar cases or future government actions concerning technology firms.

A judge has ruled that the Trump administration's blacklisting of Anthropic was illegal. This decision, reported by nytimes.com, indicates a legal challenge to the previous administration's actions. The ruling suggests that the blacklisting lacked proper legal justification, potentially impacting similar cases or future government actions regarding technology companies. The details of the ruling are available through court documents.

Over 100 tech companies, including OpenAI, Anthropic, Google, and Microsoft, have signed an open letter calling for collaboration between the private and public sectors to defend against AI-related cyber threats. These companies, while developing advanced AI models, are also offering programs like OpenAI’s Daybreak, Anthropic’s Mythos, and Microsoft’s Perception to use frontier AI for defensive purposes, highlighting their complex position in the evolving AI landscape.

A Bocconi University experiment with over 1,000 first-year undergraduates explored the impact of ChatGPT and critical thinking training on practical assignments. Students developed marketing proposals for a university merchandise store. The findings indicate that AI, specifically ChatGPT, helped students improve the quality of their answers, while critical thinking training broadened their perspectives. This suggests a complementary role for both AI and critical thinking training in preparing students for future challenges.

Tencent Holdings Ltd. has released Hy4 Preview, a foundational model with 770B parameters and a 1M context window. According to Tencent, internal tests indicate that Hy4 Preview outperforms rival models from Z.AI Co. and Moonshot AI. This new open model represents a significant development in the field of large language models, showcasing Tencent's advancements in AI technology and its competitive standing against other major players in the industry.

GLM-5.3, an open-weight model, shares its base with GLM-5.2, with all improvements stemming from post-training. It demonstrates enhanced performance in complex coding and long-horizon tasks. Key metrics show significant gains: Toolathlon Verified at 78.0, AutomationBench (v1.0.6) at 48.2, HLE w/ Tools at 28.6, and GDPval-AA v2 at 1769. The model's development is attributed to the GLM-5-Team and numerous contributors, as detailed in the paper "GLM-5: from Vibe Coding to Agentic Engineering."

Terminal-Bench-Science evaluates AI agents on scientific research workflows, with scientists setting the standards. Performance metrics for Terminal-Bench 2.1, Terminal-Bench 3.0, and Terminal-Bench-Science 0.1 show varying resolution rates and costs for models like GPT-5.6 Luna, Kimi K3, and Claude Opus 5. GPT-5.6 Sol and Claude Fable 5 demonstrate different trade-offs between cost and token usage. Development for Terminal-Bench-Science 0.2 is ongoing, with a pull request deadline of October 5, 2026, inviting researchers to contribute new tasks.

An Anthropic researcher has provided an early glimpse into the practical application of self-improving AI, a growing focus for new AI labs. This development involves training AI models using other AI models, a method gaining traction in the field. Russell Brandom, a tech industry reporter since 2012, covered this emerging technology for techcrunch.com, offering insights into what this advanced AI training could entail.

Sandhya Devanathan, Meta's India and Southeast Asia vice president, is departing the social media giant to join OpenAI, the creator of ChatGPT. This executive change comes as Meta faces increasing scrutiny in India, where it has removed 160,000 accounts over six months due to signals indicating child-exploitative activity. Meta has also addressed concerns about violating ads and accounts, rejecting claims of knowingly targeting users with inappropriate interests.

Uber has seen a significant increase in AI agent requests, with weekly requests growing 9.4x since February. Despite this surge in usage, the company has managed to keep its total AI spending stable since April. This stabilization comes after Uber reportedly used up its entire 2026 AI budget during the first quarter of the year, as reported by Madison Mills for Axios.

Sandhya Devanathan, Meta's Vice President for India and Southeast Asia, is departing the social media company to join OpenAI. At OpenAI, the maker of ChatGPT, Devanathan will be responsible for overseeing consumer growth and enterprise adoption across Southeast Asia and Australia. This move marks a significant transition for Devanathan from a major social media platform to a leading AI research and deployment company.

The "AI Engineer Notebooks" offer a framework-free approach to learning the applied-LLM stack, from prompting to serving, fine-tuning, and red-team benchmarking, using a free API on Colab. It includes case studies such as building and debugging a RAG+agent customer-support assistant, comparing agent vs. pipeline for contract extraction, and developing a red-team robustness benchmark. The program culminates in a capstone project for resume building, focusing on real-world deployment and evaluation.

Anthropic has introduced a new hardware standard called MHS, enabling AI agents to control physical devices. This allows AI models like Claude to perform tasks such as calibrating lasers or operating microscopes by adjusting devices, checking results, and repeating processes. Anthropic is collaborating with partners including Amazon Web Services, Hugging Face, and Raspberry Pi during a preview period to develop safety evaluations and best practices. MHS is designed to be an open-source, agent-agnostic standard, which could significantly accelerate technological advancements.

OpenAI and Thailand's Ministry of Higher Education, Science, Research and Innovation (MHESI) have launched an accelerator in Bangkok to support Thai AI startups. The program aims to transform prototypes into market-ready products, fostering growth and practical applications. The first cohort includes Curico, which plans to pilot its platform in Bangkok Metropolitan Administration childcare centers, train teachers, test AI-assisted grading at Chulalongkorn University, and expand its home learning platform. This collaboration seeks to establish a sustainable model for Thai founders to develop trustworthy AI products for global reach.

Owner, a company specializing in AI agents for managing restaurant websites, marketing, and other operations, has successfully raised $240M in a Series D funding round. This investment, led by Goldman Sachs Alternatives, values the company at $2.3B. The capital infusion will be utilized to further develop AI agents specifically designed for independent restaurants, enhancing their digital presence and operational efficiency.

Open-weight AI companies are becoming prime acquisition targets, with several high-profile deals recently reported. Nvidia is rumored to be acquiring Hugging Face, a platform for sharing open-weight AI models, for $13 billion. This follows Nvidia's $6 billion agreement with open-weight model builder Poolside. Additionally, Stripe recently acquired OpenRouter, a leading provider of open-weight models to businesses, for over $7 billion, highlighting a trend of significant investment in this sector.

Anthropic has launched a research preview of the Model Hardware Standard (MHS), a shared specification designed to enable AI agents to safely operate physical devices. Developed in collaboration with HHMI Janelia Research Campus, MHS allows AI agents to control various lab and manufacturing instruments, such as microscopes and robotic arms, for tasks like drug discovery and laser calibration. The MHS driver helps agents understand new devices by incorporating natural language tags for machine characteristics, generating a reference file with operational details, adjustable parameters, and safety limits.

Google announced Gemini Omni 1.1 Flash, a new model available through its genai client. This model supports interactions where users can continue a scene, as demonstrated by the input `{"type": "text", "text": "Continue the scene."}`. It also allows specifying response formats, such as `"resolution": "360p"`. The announcement, dated August 27, 2026, indicates that user information will be handled according to Google's privacy policy, with an opt-out option available.

Conduct is an open-source tool providing runtime governance for AI agents, enforcing policies across LLM calls, shell tools, and AI sessions. It includes a Router (proxy), Compliance packs, Canvas UI, Playbook DSL loader, and a Playbook library with 22 pre-built playbooks. Conduct ships with over 20 compliance packs, including OWASP, SOC 2 CC7.3, HIPAA §164.312, PCI DSS 4.0, EU AI Act Art. 15/16, NIST AI RMF, ISO 42001, and framework-specific packs for Python, Node, and Terraform.

Microsoft co-founder Bill Gates stated on Wednesday that artificial intelligence requires substantial limitations to prevent its potential harm to humans from outweighing any benefits. He discussed how AI could either reduce or exacerbate inequality, outlined three key risks associated with AI, and explored its potential impact on human pride and relationships. Gates emphasized that AI is distinct from past technologies, necessitating careful consideration and regulation.

A federal judge has blocked the Pentagon's attempt to blacklist Anthropic, ruling that the Trump administration's designation of the company as a national security risk was unconstitutional retaliation. US district judge Rita Lin vacated Defense Secretary Pete Hegseth's February 27 decision to label Anthropic a "supply-chain risk," which would have made the company ineligible for federal contracts. Lin also lifted a punitive measure preventing military contractors from doing business with Anthropic, calling Hegseth's actions "arbitrary, capricious, an abuse of discretion, and otherwise not in accordance with law."

A federal judge ruled that the Trump administration's blacklisting of Anthropic was illegal, revoking the government's directive to ban the use of the company's AI technology. The court found the Trump administration's stated reasons "untenable" and noted that federal defendants had abandoned their risk assessment. The judge clarified that Anthropic cannot backdoor its technology, and as the defendants admitted, its technology poses no greater national security risk than any other "black box" AI model.

A U.S. judge has blocked the Pentagon from blacklisting Anthropic, ruling that its designation as a supply-chain risk was "illegal and baseless." This decision marks the latest development in the Claude maker's ongoing dispute with the military regarding AI safety on the battlefield. The judge's ruling on Thursday prevents the Pentagon from proceeding with its blacklisting of Anthropic.