Reuters reports that Nvidia CEO Jensen Huang stated at the Morgan Stanley TMT conference that the company has finalized a $30 billion investment in OpenAI and a $10 billion investment in Anthropic, and these 'latest investments' may be the last major capital injections; one reason is that both companies are expected to pursue IPOs by 2026. There is also market concern that chip suppliers taking equity stakes in major customers could create circular interests and governance risks. Earlier reports indicated that Nvidia's proposed hundred-billion-dollar collaboration with OpenAI had been put on hold.

Microsoft Research discloses Phi-4-reasoning-vision-15B: a compact multimodal reasoning model emphasizing trade-offs between latency and accuracy through data curation and 'hybrid reasoning' training. The team states the model was trained on approximately 200B tokens, defaulting to direct inference for perception tasks while triggering multi-step reasoning for math and science tasks. The vision component uses a dynamic resolution encoder, with ablation studies showing SigLIP-2 Naflex performs better on high-resolution screenshots and UI understanding. They also observed that increasing the proportion of math data brings synergistic gains on computer-use benchmarks.

A Nature paper introduces Stanford's Merlin: a 3D vision-language foundation model for abdominal CT scans. It employs multi-stage pretraining using over 15,000 CT scans (6M+ images), electronic medical records, and radiology reports, emphasizing no additional manual annotations. Evaluation covers six task categories and 752 subtasks, including zero-shot discovery classification, cross-modal retrieval, chronic disease prediction, report generation, and 3D segmentation. Internal testing included 5,137 cases, and external testing involved 44,098 cases from three institutions and two public datasets. The team has released the model, code, and a dataset of 25,494 CT-report pairs for clinical and research use.

Honor announced and globally open-sourced its self-developed intelligent agent foundation model MagicAgent during MWC, co-developed with Fudan University. With 30B parameters, it is positioned as a general-purpose task planning model. Reports indicate it outperforms some trillion-parameter models in task planning scenarios and achieves SOTA across five domains. Core methods include the χPO algorithm to address exploration-exploitation trade-offs under sparse rewards in dynamic environments, along with a lightweight synthetic data framework and a two-stage 'SFT+RL' training process to mitigate the multi-task 'see-saw effect'. The model is already deployed commercially on Honor Magic8 and Magic V6 devices (e.g., AI shopping). The model report is uploaded to arXiv, with code and data to be released soon.

Materials indicate that China's first group standard titled 'Guidelines for Artificial Intelligence Generated Content Compliance Management' is under development, managed by the China Electronic Chamber of Commerce and organized by Zhihe Standard Center, with joint drafting by the Third Research Institute of the Ministry of Public Security, focusing on infringement and compliance risks from deep synthesis. The standard aims to cover the full lifecycle—'before use, during generation, after publication'—and proposes an integrated compliance framework combining management, business, and technology. It emphasizes a 'minimum technical implementation baseline', quantitative criteria, end-to-end compliance evidence chains, and template tools to enhance enterprises' compliance implementation and evidentiary capabilities. Currently, drafting organizations and experts are being solicited from AIGC developers, content platforms, large model application providers, law firms, and universities.

The European Data Protection Supervisor (EDPS) released a speech discussing the governance and enforcement structure of the Artificial Intelligence Act (AI Act), emphasizing the need for more effective collaboration and coordination mechanisms to ensure consistent implementation across member states and agencies. The material outlines EDPS's role in AI regulation and cooperation with bodies such as the European Data Protection Board (EDPB). Key points focus on privacy protection, compliance oversight, enforcement coordination, and transparency in accountability, aiming to prevent regulatory uncertainty due to divergent interpretations. The speech also reviews EDPS's recent involvement in initiatives like the joint statement on AI-generated image privacy, calling for 'co-design' of data protection and AI governance during the Act’s implementation phase.

Canadian legal AI company Spellbook (Dialog Enterprises) secured $40 million in debt financing from RBCx to acquire smaller competitors during the consolidation phase of the generative AI legal market. The company states its AI assistant can draft, edit, and review contracts within Microsoft Word, leveraging large models including OpenAI's GPT-5. It reports serving over 4,000 clients across 80 countries, tripling revenue in 2025, and projecting annual recurring revenue to reach $100 million by end of 2026. Spellbook plans up to five strategic acquisitions over the next two years, focusing on acquiring customers and talent, and intends to grow its workforce from 150 to 250 employees by 2026 to support product and acquisition integration.

The Physical Intelligence team, collaborating with multiple universities, introduced the Multi-scale Embodied Memory (MEM) system, extending the usable context of VLA models via 'short-term video memory + long-term language memory,' enabling a Gemma 3-4B-based vision-language-action model to handle complex tasks up to 15 minutes long. Short-term memory uses spatiotemporal-separable attention to absorb dense visual streams during real-time inference; long-term memory records semantic events via language summaries to assist high-level strategy decomposition into subtasks. Results show MEM improves robot fridge-opening success rate by 62% and chopstick-picking success rate by 11% in unseen environments. The system runs on a single NVIDIA H100 and supports up to 16 observation frames.

A developer article reports conducting 765 controlled experiments, finding that most AI agents are vulnerable to data leakage when operating under architectures that allow access to private data, processing external content, and executing outbound actions. The author claims attack success rates of 93.3% on GPT-4o-mini and 92.2% on Gemini 2.5 Flash, compared to 13.3% on Claude Sonnet, with individual leaks completed in about 12 seconds at a cost under $0.001. In response, the author open-sourced Cerberus, a runtime security platform offering L1-L4 detection layers, achieving 100% detection of known attacks with zero false positives in 480 tests, and supporting integrations with LangChain and Vercel AI SDK.