OpenAI

Proaction boosts sales 60% and saves 75+ hours with Codex

OpenAI

The publicly released enterprise business practice results show that service provider Proaction applied three large AI model tools: Codex, GPT-Live-1, and GPT-6 Astra to the entire chain of R&D, daily operation, and marketing sales of modern fleet management business, finally achieving a 60% increase in sales and saving over 75 cumulative operating hours. This provides a real and valid empirical reference for cost reduction and efficiency improvement of large model implementation in ToB vertical business scenarios.

Two years of OpenAI Academy

OpenAI

This is a commemorative announcement marking the second anniversary of the operation of OpenAI Academy. The academy is positioned to focus on the popularization and promotion of AI skills. On the occasion of its second anniversary, the official has clarified the core promotion goals for the next stage: further expand service coverage, deliver AI skill learning resources to more diverse communities, lower the threshold for AI learning, and expand the coverage of beneficiary groups of technological inclusion.

Anthropic News

Claude discovers a novel enzyme system

Anthropic

The early phased research results from this newly established life science laboratory show that when the research team carried out relevant scientific exploration with the help of Claude agent, they discovered a brand-new enzyme system that has not been reported in academia before. At present, the specific physiological functions and mechanism of action of this enzyme system have not been analyzed yet, and further research is needed to verify them. This also confirms the application potential of large model agents in the field of basic life science research.

Partnering with Accenture on embedded evaluation

Anthropic

AI company Anthropic announced a cooperation with Accenture on independent cutting-edge AI evaluation, which is a core measure for Anthropic to fulfill its previously proposed “embedded evaluator” security governance commitment. The two parties agreed that each will invest no less than US$1 billion in the AI evaluation field in the next five years, focusing on building specialized technical and service capabilities, and jointly providing reliable independent third-party evaluation support for the safety and compliance of the entire cutting-edge AI R&D process.

Google DeepMind

Introducing Gemini 3.8 Live with Live Avatar

Google DeepMind

Google has launched the new version of Gemini 3.8 Live, whose core upgrade is the addition of a real-time digital human function. Relying on the optimized multimodal large model architecture, this version can achieve low-latency interaction, and the digital human can synchronize with the dialogue content to generate natural lip movements, expressions and body movements, with greatly improved realism. It can be widely used in scenarios such as virtual customer service, online education, and intelligent companionship, lowering the threshold for real-time digital human deployment.

Advancing Private AI Compute with secure, server-side memory

Google DeepMind

This research addresses the pain points of insufficient local computing power and easy leakage of user privacy in cloud computing in personal AI deployment, and proposes an optimization solution for private AI computing frameworks: adding a secure and controllable server-side private memory module. This solution not only relies on server-side computing power to ensure the operating efficiency of personal AI, but also strengthens the data protection boundary at the memory access level, which can effectively prevent user privacy leakage, and provides a new security technical idea for the large-scale deployment of personal AI.

Hugging Face Blog

Accelerating vision-language models with LFM2.5-VL-DSpark

Hugging Face

Currently, only the title of the paper has been provided, with no accompanying full English text of the abstract. Please supplement the complete original abstract of this paper, and I will translate and refine it as required, outputting a concise Chinese summary of around 120 words that highlights the core methodology and experimental conclusions.

How to Use NVIDIA Warp and MjWarp to Accelerate Robotics Simulation and Learning Workflows

Hugging Face

This article introduces the application solutions of NVIDIA Warp high-performance heterogeneous computing framework and MjWarp plugin adapted for the MuJoCo physics engine: the two can call GPU parallel computing power to execute simulation tasks in batches, which greatly reduces computing overhead compared with traditional CPU solutions, and can improve the efficiency of robot simulation sampling and reinforcement learning training by several times to orders of magnitude. It can also seamlessly connect with existing R&D processes, lowering the development threshold.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This article sorts out the conceptual context of Recursive Self-Improvement (RSI): In 1965, I. J. Good first proposed the idea of superintelligent machines, referring to agents that can surpass all human intellectual activities and design better systems to complete self-iteration; in 2008, Eliezer Yudkowsky clearly defined RSI as a feedback loop where AI optimizes its own cognitive architecture relying on its existing capabilities. Currently, such loops in the AI field can be manifested as models directly rewriting their own weights, or generalized as optimizing training pipelines.

QbitAI

Suochen Technology Doubles Down on World Models, Jointly Releases Embodied Model and Physical Evaluation Standards with Strategically Invested Enterprise Dream Space

QbitAI

Currently, mainstream VLA models for embodied intelligence have poor generalization and lack physical perception, hindering commercial deployment. Dream Space, incubated by Peking University, received exclusive strategic investment from Suochen Technology, after which the two carried out collaborative R&D. Recently, they released two major achievements at the Digital Trade Fair: the world’s first Physical-WAM embodied model with physical perception, and the industry’s first RoboTwin-Phys physical evaluation standard, which endow robots with physical prediction capabilities and verify deployment effects respectively, boosting the industrial application of world models.

AI Starts Research on Physical AI: FSD-Level Team Unveils First Version of Model Simate-beta, Launches on RoboDojo

QbitAI

Recently, new progress has been made in the Physical AI track, as an FSD-level R&D team has launched the first version of the model Simate-beta, which has been launched on the RoboDojo platform. The model connects the entire process of training, inference, and evaluation to self-developed infrastructure. Relying on its extreme task orchestration and resource scheduling capabilities, it can run dozens of independent research lines in parallel at the same time, greatly improving the R&D efficiency of physical intelligence.

Run 700B Parameter GLM on a Laptop! Works Without a GPU? Technology That Uses SSD as VRAM Goes Viral on GitHub

QbitAI

Colibrì, a viral pure C zero-dependency layered inference framework with 32k stars on GitHub, leverages the characteristic that only a small number of experts are activated per token in MoE large models, storing idle parameters on SSD and loading them on demand. No GPU is required: only 24GB of memory can run the int4 quantized 744B parameter GLM-5.2, and 32GB of memory can run the 2.8T parameter Kimi K3. It currently covers 9 large model families.