Adobe Sued Over AI Voice Training — The Wave of Litigation Grows

Adobe has become the latest target of litigation over AI voice training. According to Bloomberg Law, claims that AI companies have been using voice data for training without authorization keep piling up, and Adobe is positioned as the most recent target in this 'wave' of lawsuits. Legal uncertainty around AI copyright and data use continues to expand.

Git Isn't Ready for the AI Coding Wave

As The Register points out, Git is struggling to keep up with the massive volume of code changes generated by AI coding agents. With PRs ballooning in size and review effectiveness declining, there is growing debate over the need to rethink version control itself. As AI agents move into full-scale deployment, the limitations of the tools underpinning development workflows are starting to surface.

Linux Kernel Adds New Guidelines for AI Bug Reporting

The Linux Kernel 7.1 documentation now includes a definition of security bugs along with guidelines for responsible AI-related reporting. According to Phoronix, the addition spells out the reporting flow for vulnerabilities discovered using AI tools, showing that open source communities are making progress on codifying norms for AI use.

Orthrus: Up to 7.8x Faster Inference with Qwen3-8B

Orthrus-Qwen3-8B is drawing attention in Reddit's LocalLLaMA community. Using a two-stage approach that injects a trainable diffusion attention module into a frozen Qwen3-8B backbone and verifies with an AR (autoregressive) head, it reportedly accelerates inference by up to 7.8x while theoretically keeping the output distribution identical to the original model. Unlike existing Speculative Decoding, it requires no external drafter model and incurs zero TTFT (Time-To-First-Token) penalty. On the MATH-500 benchmark, it achieved an acceptance length of 11.7, far surpassing DFlash's 7.9 and EAGLE-3's 3.5.

AllenAI Continues Rolling Out MolmoAct2, a Robot-Control VLA Model

AllenAI (Allen Institute for AI) has been releasing a steady stream of fine-tuned versions of MolmoAct2, a 5B-parameter Vision-Language-Action model. Models for diverse robotics datasets — LIBERO (general robotics), DROID (interactive manipulation), and BimanualYAM (bimanual absolute joint-pose control) — are available on Hugging Face, and the training datasets, training software source code, and technical papers are all provided as open source.

Gemma4 26B MoE Reaches Practical Speeds on MLX

A community effort to run Gemma4 26B MoE on a MacBook Air M5 using the MLX framework and TurboQuant has been reported. It can handle 4 parallel batches at 128K context, and at 8K context it reportedly outperformed llama.cpp across the board in prompt processing, generation speed, and memory usage. Custom kernels for SWA layers unlock the real memory savings of 2-bit quantization.

Snapdragon X2 Laptop Processes 200K Documents with RAG

A demo has been released showing a laptop powered by Qualcomm's latest Snapdragon X2 chip indexing roughly 200,000 files and running RAG (Retrieval-Augmented Generation) over them. Leveraging VecML's integrated AI database platform, it achieves low-cost retrieval with about 1,200 tokens and low-memory operation — a step closer to running serious RAG workloads on edge devices.

---