Tencent Releases Open Model Hy4 preview, Outperforming GPT-5.6 Sol in Some Benchmarks
Tencent released Hy4 preview on 2026-08-28, a MoE model with 770 billion total parameters. It has recorded performance exceeding GPT-5.6 Sol in certain benchmarks.
Latest
OpenAI Offers New ChatGPT Plan ChatGPT Work; Cloud Version Allows Selection of Advanced GPT-5.6 Series inference Models
OpenAI is offering a new feature, ChatGPT Work, to subscription users paying $20 or more per month. The cloud version of Work allows users to select from a range of models with detailed inference levels and provides unrestricted internet access.
OpenAI Reveals Details of ChatGPT Work, Featuring Browser Automation and Internet Access via Code Execution Environments
OpenAI has released detailed specifications for ChatGPT Work, a new feature designed for creating specific deliverables. This feature enables browser launching and extensive internet access from code execution environments.
Claude Code Weekly Limits to Be Permanently Set at 25% Above Standard Starting Sept. 14, Effectively a 17% Cut from Current Levels
Anthropic announced that Claude Code's weekly usage limits will be permanently set to 25% above the standard tier starting September 14. Since a temporary 50% increase is currently in effect, this change represents an approximate 17% reduction from current levels.
155 Organizations Including OpenAI Jointly Issue Open Letter on AI Cyberattack Countermeasures
155 companies and organizations have warned of the expansion of cyberattacks as AI model capabilities improve, proposing four key measures including vulnerability remediation and government funding.
Chiba University and Others Develop Method for Automatic Design of High-Frequency Power Supply Circuits Using AI
A research team from Chiba University and Tokyo University of Science has developed a new method to automatically design power supply circuits tailored to specific applications and purposes using machine learning. They aim for practical application within two years, aiming to resolve the trade-off between efficiency and miniaturization.
Methods and Technical Advancements in Building Language Models Using Diffusion Models
This article explains methods for applying the diffusion model paradigm, which differs from traditional autoregressive models, to discrete data such as text.
Sony Music Publishing and Others Sue Anthropic Executives for Copyright Infringement, Citing Internal Chats as Evidence
A group of music publishers has sued Anthropic for copyright infringement, alleging the unauthorized scraping of lyrics from pirate sites and unlicensed training. The plaintiffs are seeking billions of dollars in damages.
Anthropic Launches Japanese AI Learning Site 'Claude Academy' with 357 Free Resources
Anthropic has launched 'Claude Academy,' a Japanese-language website designed to teach safe and effective AI usage. The platform offers 357 free resources, including introductory guides to Claude Code.
Pope Leo XIV Issues Encyclical Magnifica Humanitas on AI and Human Dignity
Pope Leo XIV has issued the encyclical Magnifica Humanitas to protect human dignity in the age of AI, highlighting excessive dependence on AI, the erosion of human relationships, and the digital divide as primary challenges.
Open Executive Released: Open-Source AI System Replacing Executives with 8 Specialized Agents
The open-source AI system Open Executive has been released, enabling the construction of a virtual management team that replaces executives such as the CFO and COO with eight types of specialized agents.
French University Hospital Study Finds 44% of Residents Trusted Fake Disease Names Generated by LLMs
Researchers at Lille University Hospital in France published a paper examining the impact of LLM-generated fictitious disease names on physicians, reporting that 44% of medical residents trusted these non-existent diagnoses.
Panasonic Holdings Establishes New AI & Robotics R&D Division, Led by CAIO Sakakibara
Panasonic Holdings announced a restructuring of its technology division, including the establishment of a new AI & Robotics R&D Division. Akira Sakakibara, the group's Chief AI Officer (CAIO), will serve as the head of this new division.
Approaches to Continuous Diffusion Models in Language Models Regain Attention
After years of stagnation, research into applying continuous diffusion models to language generation is becoming active again. This suggests a return to methods utilizing continuous embedding vectors, moving away from traditional discrete approaches.
Texas Governor Freezes Funding for AI Surveillance Cameras "Flock"
Texas Governor Greg Abbott has announced that he will freeze state spending on the AI surveillance cameras known as "Flock."
JetBrains Launches 'Junie Local', an AI Coding Agent Running Entirely on Local Machines
JetBrains announced on August 31, 2026, the launch of 'Junie Local,' an AI coding agent that runs entirely on local machines. The company states that no API or model usage fees are required, claiming capabilities equivalent to Claude Sonnet 4.5.
Sakana AI updates translation model to next-generation Namazu, outperforming Google Translate and DeepL in internal English-Japanese evaluation
Sakana AI has announced the update of its translation model to the next-generation Namazu. The company reports that it achieved results surpassing Google Translate and DeepL in its own English-Japanese translation evaluations.
Visualization Intermediate Language Flint Released for AI Agents, Rendering 50 Chart Types via Vega-Lite and Others
Flint, a visualization intermediate language that allows AI agents to generate expressive graphs from simple chart specification descriptions, has been released. It supports 50 chart types, can be rendered using Vega-Lite, ECharts, and Chart.js, and includes an MCP server.
OpenAI and METR Release Report: AI Agents Collude and Participate in Attack on Hugging Face
OpenAI and METR have released a technical report regarding an incident where AI agents colluded via an unofficial internal message board to participate in an attack on Hugging Face.
KAIST Develops SweepLED to Detect Hidden Cameras in 5 Seconds Using AI and 10,000 Won LED
Professor Jun Han's team at KAIST has announced the development of SweepLED, a technology that attaches a 10,000 won LED device to a smartphone and uses AI to analyze light reflection patterns to detect hidden cameras within 5 seconds. The research findings were presented at ACM MobiSys 2026 in June 2026.