Models & Labs
Zhipu AI's GLM-5.2 rivals Mythos in cybersecurity
The Verge AIhigh confidence
Why it matters
- →GLM-5.2's capabilities in cybersecurity highlight China's progress in AI development.
- →The open-weight nature of the model allows for widespread use and potential misuse.
- →This development raises national security concerns for the US government.
Zhipu AI, a Chinese company, has released its open-weight model GLM-5.2, which reportedly matches the cybersecurity capabilities of Mythos, a model from Anthropic. While GLM-5.2 lags behind in general AI tasks compared to models from Anthropic and OpenAI, its proficiency in bug-finding is notable. This advancement is concerning for the US government, which sees such models as potential national security threats. The open-weight nature of GLM-5.2 allows it to be run on readily available hardware, increasing its accessibility and potential for misuse.
Read originalMore from The Verge AI
© The Verge AIMarket & Regulationmusic
Tidal to Demonetize AI-Generated Music
Tidal is taking a decisive step by demonetizing tracks identified as fully AI-generated, effective immediately. While these tracks won't be banned, they will be marked with a specific icon starting July 15th to inform listeners. This initiative aims to ensure that royalties are directed towards music created by humans, addressing concerns about fair compensation. As the music industry grapples with the rise of AI, Tidal's approach underscores the need for reliable detection tools and accurate labeling. The platform plans to expand its labeling to include music that is substantially AI-generated as technology improves, urging content distributors to properly label AI-generated music.
The Verge AI
© The Verge AIModels & Labscoding
OpenAI Teases New Hardware for Codex
OpenAI is stepping into the hardware space with a new device tailored for its AI-powered coding tool, Codex. In collaboration with Work Louder, known for their mechanical keyboards and macro pads, OpenAI is set to launch a device that promises to enhance Codex shortcuts. The teaser suggests a device similar to Work Louder's Creator Micro 2, which features customizable mechanical switches and a joystick. This move could streamline coding workflows by integrating physical controls with AI capabilities, marking a novel intersection of hardware and AI in coding environments.
The Verge AI
© The Verge AIMarket & Regulationbusiness
Lawmakers Propose Ban on AI Health Data Sales
Lawmakers are taking a decisive step towards safeguarding personal data by proposing a bill to prevent AI companies from selling health and location data to brokers. This initiative, led by Senator Elizabeth Warren and Representative Mary Gay Scanlon, seeks to update existing legislation to better fit the AI-driven landscape. As AI tools like ChatGPT and Claude increasingly manage sensitive health information, the bill aims to ensure that such data is not commercially exploited. If enacted, it would grant the FTC authority to enforce these protections, allocating $1 billion for this purpose over the next decade. This proposal represents a crucial move towards establishing a robust data privacy framework as AI becomes more embedded in healthcare.
The Verge AI
More in Models & Labs
Models & Labsmodels
vLLM v0.24.0 Release Enhances Model Support
The vLLM v0.24.0 release marks a significant update with extensive contributions from 256 developers, introducing support for new models like MiniMax-M3 and DiffusionGemma. This version enhances performance with optimizations such as the FlashInfer sparse index cache and improved throughput for DeepSeek-V4. The update also expands the Model Runner V2 capabilities, supporting quantized models by default and integrating GraniteMoE. These advancements make vLLM more robust and versatile, offering developers improved tools for model deployment and performance tuning.
vLLM Releases
Models & Labsmodels
Llama.cpp b9833 Release Enhances MiniCPM5 Parser
The latest b9833 release of llama.cpp focuses on refining the MiniCPM5 parser, addressing several technical aspects to improve its functionality. This update includes the addition of a new tool call parser, refactoring of the PEG parser, and adjustments to the Jinja min/max API for better compatibility with Jinja2. The release also reverts some shared mapper changes to maintain strict JSON parsing for tool-call arguments. These enhancements aim to streamline the parsing process, ensuring more reliable and efficient handling of XML tool calls and grammar triggers.
llama.cpp Releases
Models & Labsmodels
llama.cpp b9835 Release Expands Platform Support
The latest b9835 release of llama.cpp continues its trend of broadening platform compatibility, though without major new features. Notably, the release includes support for ROCm 7.2 on Ubuntu x64, which is significant for AMD GPU users seeking alternatives to NVIDIA's CUDA. The update also maintains a wide array of builds across macOS, Linux, Windows, and openEuler, ensuring developers have the flexibility to deploy on diverse systems. While the release doesn't introduce groundbreaking changes, it solidifies llama.cpp's position as a versatile tool for AI inference across multiple environments.
llama.cpp Releases