xAI filed a federal suit in the Central District of California naming Attorney General Rob Bonta, asking the court to declare the Generative AI Training Data Transparency Act unconstitutional. AB 2013 took effect January 1 and requires developers of generative systems available in California to publish dataset documentation across twelve categories — sources, data types, sizes, copyright status, whether data was purchased or licensed, and more.

The 55-page complaint calls it an “unconstitutional trade-secrets-destroying disclosure regime” that compels speech and takes property without compensation, arguing dataset composition is a core competitive secret.

Alibaba's Tongyi-MAI team released MAI-UI, foundation GUI agents at 2B, 8B, 32B and a sparse 235B-A22B, built on a Qwen3-VL backbone. The 2B and 8B weights are Apache 2.0 on Hugging Face.

Reported results: 76.7% on AndroidWorld and 73.5% on ScreenSpot-Pro. The engineering is the interesting part — a device-cloud split that routes execution by task state, reported 33% better on-device performance and 40%+ fewer cloud calls, plus native MCP tool calls and an online RL setup where scaling parallel environments from 32 to 512 added 5.2 points.

Tencent Hunyuan released HY-Motion 1.0, text-to-3D human motion models built on a Diffusion Transformer with flow matching — by their account the first time DiT-based text-to-motion has been scaled to the billion-parameter level.

The recipe is a fully documented three-stage stack: pretraining on 3,000+ hours of motion data, fine-tuning on 400 curated hours, then RLHF with reward models to tighten instruction following. Inference code and weights are on Hugging Face with a demo Space.

Meta said it would acquire Manus, the Singapore-based general agent company that went viral in early 2025. Terms were not disclosed, but a source with direct knowledge put the deal at $2–3 billion.

Manus says it keeps running and selling its subscription product from Singapore while being integrated into Meta's consumer and business surfaces, including Meta AI. Its own numbers: more than 147 trillion tokens processed and over 80 million virtual computers powered.

NPR reported that AI data centers have soaked up so much memory capacity that RAM supply trails demand, and the imbalance is about to show up in the price of computers, phones and consoles. TrendForce's Avril Wu told NPR demand exceeds supply by about 10%, buyers paid roughly 50% more for common DRAM this quarter, and rush orders cost two to three times more — with another ~40% increase expected next quarter.

Micron posted better-than-expected earnings on the higher prices and said supply will remain “substantially short of the demand for the foreseeable future.”