OpenAI

Our decision on Cursor following its acquisition by SpaceX

OpenAI

This statement is released by the partner that previously provided OpenAI LLM support for the code editor Cursor. The core decision is to terminate the previous model supply service contract between the two parties, with the direct trigger being that Cursor has been acquired by SpaceX. The statement currently does not disclose the deep-seated considerations for terminating the cooperation, the subsequent model replacement plan for Cursor, nor the actual impact of this change on ordinary users.

Supporting Thailand’s next generation of AI startups

OpenAI

This project is jointly launched by OpenAI and Thailand’s Ministry of Higher Education, Science, Research and Innovation, with the core goal of nurturing Thailand’s next-generation AI startups. The program is an 8-week special acceleration plan covering 10 startups in three vertical fields: healthcare, wellness and elderly care, and education. It will provide support to them, promote the implementation of existing AI technology prototypes into reliable and mature products, and help build Thailand’s AI industry ecosystem.

Anthropic News

Previewing the Model Hardware Standard

Anthropic

Recently, Anthropic launched the research preview of the Model Hardware Standard (MHS), which is initially open to research laboratories and high-end manufacturers. This standard is a general shared specification for AI agents to safely control physical devices, which will unify the interaction interfaces and security guidelines between agents and physical hardware, provide a unified collaboration foundation for embodied AI implementation and intelligent applications in industrial scenarios, and facilitate coordinated exploration by industry, academia and research institutions.

How Claude’s text watermarking works

Anthropic

Anthropic discloses that future Claude series LLMs will generate watermarked text by default. This technology can trace and determine whether text is generated by Claude, and this move is a compliance action taken in conjunction with multiple leading AI vendors to implement the requirements of the EU AI Act. This article focuses on responding to three types of public concerns: the implementation logic of the watermark technology, whether the watermark affects output quality, and the motivation for rolling out this feature.

Google DeepMind

Gemini Omni 1.1 Flash lets you build with more control

Google DeepMind

Google’s newly released Gemini Omni 1.1 Flash is a lightweight multimodal LLM for developers, focusing on higher controllability. It adds capabilities such as fine-grained output format constraints, custom behavior anchoring, and adjustable multimodal output logic, supporting developers to flexibly align with business needs. Compared with the previous generation, the development and adaptation cost of this model is reduced by about 30%, output compliance and custom matching degree are significantly improved, making it suitable for application integration scenarios with low latency and high customization requirements.

Piloting the world’s first double-blind AI evaluations

Google DeepMind

This article introduces the world’s first double-blind AI evaluation pilot. The evaluation adopts an innovative double-blind paradigm: evaluators do not know in advance whether the evaluation subject is AI or human, and the participating AI is unaware throughout the process and will not adjust its output targeted, which can effectively avoid the observer bias and targeted cheating problems of traditional AI evaluations. The pilot verifies that this paradigm is highly feasible, can greatly improve the objectivity of AI capability evaluation, and provides reliable new support for AI capability grading and compliance supervision.

Hugging Face Blog

The Open ASR Leaderboard Adds Its First Global South Language

Hugging Face

Previously, the open-source Automatic Speech Recognition (ASR) public leaderboard only covered high-resource languages in Europe and the United States. This research includes Swahili, a representative language of the Global South, for the first time. Methodologically, it optimizes the evaluation benchmark for low-resource languages and supplements annotated datasets covering multiple scenarios and dialects. Evaluation shows that the accuracy of mainstream open-source ASR for Swahili is 42% lower than that for English, exposing the technical gap for low-resource languages and pointing out the direction for multilingual ASR research and development.

Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers

Hugging Face

This article addresses the pain points of coarse semantic representation granularity and poor performance in complex semantic scenarios of single-vector embeddings. It optimizes the training and fine-tuning process of multi-vector embeddings based on the Sentence Transformers framework, introduces block-level semantic alignment and hard negative sample sampling mechanisms. The accuracy on information retrieval and semantic matching tasks is improved by 12%~18% compared with the single-vector solution, with inference efficiency loss of less than 10%, which can be quickly implemented adapting to the existing ecosystem.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

The concept of Recursive Self-Improvement (RSI) can be traced back to the superintelligence hypothesis proposed by I.J. Good in 1965: such a system can surpass humans in all intellectual activities, and can also independently design better intelligent systems. In 2008, Eliezer Yudkowsky clarified that its core is the intelligent feedback loop: AI uses its existing capabilities to iteratively upgrade its own cognitive architecture. RSI in the current AI context includes both the model directly rewriting its own weights, and broadly covers the optimization of its own training process.

量子位

The culprit of locally deployed AI performing worse than the official version found: 734 dependency packages, each can be a pitfall

量子位

The core reason why locally deployed LLMs perform worse than the official version has been identified: researchers used 100k token workflow data that does not appear in public benchmarks or training sets, tested Qwen3.6-27B under the same GPU and same weight conditions, and found that differences in the inference stack (such as switching attention backends, floating-point operation precision/accumulation order, instruction set differences) will cause logit deviation. Once the deviation is sufficient to change the highest probability token, it will greatly affect the output, and even lead to tool call failure.

Mengdi Xu, who returned to China last year, becomes homeroom teacher of Tsinghua Yao Class

量子位

Tsinghua’s 2026 Yao Class admitted about 90 top students, including those recommended through competitions, national training team members recruited in the second round, Olympiad national team members, top college entrance examination performers, etc. Mengdi Xu, an assistant professor at the Institute for Interdisciplinary Information Sciences who returned from Stanford in 2025 with an interdisciplinary background (bachelor’s degree in vehicle engineering, PhD from CMU focusing on robot learning and AI safety), became the homeroom teacher of Yao Class freshmen less than a year after joining the faculty, with the welcome sharing theme of “Building Next-Generation Intelligence”.

Claude starts training Claude! $4 per hour, outperforms human researchers paid $150 per hour

量子位

Anthropic’s newly launched AAR (Automated Alignment Researcher) system built on Claude Opus 4.8 can independently complete the entire research process of searching literature, proposing solutions, generating data, and fine-tuning models, at a cost of only $4 per hour. Tests show that it can find improvement solutions for all 10 types of AI alignment security problems, and performs better than human researchers with an hourly salary of $150 on some tasks. Weaker versions of Claude can already participate in training stronger versions, accelerating the process of AI self-iteration.