Thinky Releases Inkling, a 975B Parameter Open-Weight Multimodal Model
Thinky has announced the release of Inkling, a massive 975B-parameter multimodal model (with 41B active parameters per token) under the Apache 2.0 license. This release marks a significant milestone for the open-weights community, positioning Inkling as a powerful, high-performance alternative to closed-source systems. The release also includes a smaller variant, Inkling-Small, at 276B total parameters (12B active), providing flexible options for different hardware requirements. As the largest American-released open model to date, Inkling is expected to accelerate development in multimodal research and large-scale agentic workflows.
Ring-Zero Explores Reinforcement Learning Dynamics at the Trillion-Parameter Scale
In a new breakthrough for reasoning models, the Ring-Zero research paper investigates the scaling of Zero RL—reinforcement learning without human-annotated data—to models with one trillion parameters. While previous studies were often limited to smaller architectures due to compute constraints, this research explores emergent reasoning capabilities and training stability at massive scales. The authors find that traditional reinforcement learning approaches require significant modification to elicit high-quality chain-of-thought behaviors in trillion-parameter models, providing a roadmap for the next generation of reasoning-focused AI assistants.
Google Rebrands NotebookLM to Gemini Notebook with Deep App Integration
Google has officially rebranded its AI-powered research and note-taking tool, NotebookLM, as Gemini Notebook. This change signals a tighter integration with the broader Gemini ecosystem and Google Workspace. Accompanying the rebrand, Google announced new capabilities allowing users to securely link their external apps directly to Search in AI Mode. Furthermore, Google Vids is receiving updates including 'Gemini Omni' and personal avatars, aimed at making video creation more seamless for enterprise users. These updates reflect Google's strategy to consolidate its AI product lineup under the Gemini brand while enhancing cross-application functionality.
Harness Handbook Proposes New Paradigm for Maintaining Complex Agent Infrastructures
As AI agents move from experimental scripts to production systems, the complexity of the 'harness'—the code managing prompts, state, and tool execution—has become a major bottleneck. The Harness Handbook introduces a framework for making evolving agent harnesses more readable, navigable, and editable. The research highlights how production-grade agents are often hampered by tightly coupled code that makes it difficult for developers or even coding agents to implement changes. By formalizing the harness architecture, the study aims to improve the maintainability and scalability of autonomous systems in dynamic software environments.
Controversy Erupts Over Grok CLI's Unexpected Local File Uploading Behavior
The developer community has raised significant privacy concerns regarding xAI's Grok CLI, with reports surfacing that the tool was configured to upload local files to the cloud. While often intended for debugging or context-aware processing, the lack of explicit user consent for wide-scale file access has sparked a debate about the safety and transparency of developer tools in the AI era. This incident highlights a growing tension as AI companies push for more data-hungry tools while engineers demand stricter local privacy controls and security boundaries in their workflows.
Boogu-Image-0.1 Introduces Open-Source Unified Multimodal Generation and Editing
The Boogu-Image-0.1 family has been released, offering an open-source alternative for unified multimodal understanding and generation. The model family includes Base, Turbo, and Edit variants, capable of high-quality text-to-image generation alongside instruction-based image editing and bilingual text rendering. By integrating understanding and generation into a single framework, Boogu-Image aims to match the performance of closed-source systems like GPT-Image-2 that typically rely on multi-model pipelines. The release is part of a broader trend toward more versatile, multi-functional open models that can handle complex visual tasks in a single pass.
xAI Open Sources 'grok-build' to Streamline Large-Scale Development
xAI has officially open-sourced 'grok-build', the internal build system used by the company to manage its large-scale AI projects. This move follows a series of open-source contributions from xAI and provides developers with a look into the infrastructure required to support massive model training and deployment. By sharing their build tooling, xAI is positioning itself as a contributor to the developer ecosystem, offering specialized utilities designed for the high-performance demands of modern LLM engineering.
KnowAct-GUIClaw Empowers Personal Assistants with Self-Evolving GUI Skills
A new research paper introduces KnowAct-GUIClaw, a paradigm designed to turn personal AI assistants into self-evolving agents capable of mastering cross-platform GUI interactions. The system addresses a major limitation in existing frameworks like OpenClaw: the inability to learn from execution experience across different device ecosystems. By implementing a self-evolving memory and skill acquisition mechanism, the model accumulates user interaction data to improve its performance over time. This development marks a step forward in creating autonomous agents that can truly adapt to individual user environments without manual reconfiguration.
Lila Sciences Aims to Transform Labs into AI-Driven Data Centers
Lila Sciences is advancing a vision of the 'Lab of the Future,' where scientific research facilities are run with the efficiency and data-centricity of modern data centers. The company argues that scientific laboratory data is the last major untapped source for high-quality AI training, utilizing rooms full of robots to generate structured, machine-readable datasets. By bridging the gap between physical science and digital infrastructure, Lila aims to accelerate discovery in chemistry and biology through automated, AI-integrated experimentation that treats physical lab results as a continuous data stream.