NVIDIA and LG Group Announce Comprehensive Partnership to Build AI Factories

NVIDIA and South Korea's LG Group have partnered to build large-scale AI factories spanning physical AI, robotics, autonomous driving, data center infrastructure, and GPU cloud services. The partnership combines NVIDIA's end-to-end AI factory platform with the LG Group's global strengths in home appliances, robotics, mobility, and data center technologies.

In **robotics**, LG Electronics is developing a home robot called "CLoiD," enabling simulation and training in physically accurate virtual environments built on NVIDIA Isaac Sim and Isaac Lab. LG also plans to bring NVIDIA Isaac GR00T (an open reasoning vision-action language model) to both its home robots and its modular robotics platform, positioning LG robots as part of the NVIDIA GR00T ecosystem.

LG Electronics further plans to develop a physical AI data factory, using synthetic data generation with the NVIDIA Cosmos world foundation model to supply high-quality training data for robotics and industrial AI.

For **AI factory infrastructure**, the partnership will push forward with building liquid-cooled AI factories aligned with the NVIDIA DSX platform. LG Energy Solution will work with NVIDIA on 800V direct current data center power solutions to meet the power demands of next-generation GPUs.

In **autonomous driving**, LG Electronics will integrate sensors, compute, and software into the NVIDIA DRIVE Hyperion architecture. AI cockpits and edge AI processing leveraging NVIDIA DRIVE AGX are also on the roadmap.

In **AI models**, LG AI Research will advance the development of "EXAONE," South Korea's flagship sovereign AI model, leveraging NVIDIA Blackwell GPUs, the NeMo framework, and Nemotron datasets. ChatEXAONE will accelerate the deployment of AI technologies across the LG Group.

The partnership shows how the AI factory ecosystem is expanding across multiple layers, from chips and power to robotics and sovereign AI.

dots.tts: A 2B-Parameter TTS Foundation Model Arrives

A technical report has been released for dots.tts, a 2B-parameter continuous autoregressive text-to-speech (TTS) foundation model.

The model combines Audio-VAE, flow matching (full-history conditioning), and reward-free self-correcting post-training, reportedly achieving state-of-the-art speech generation performance. ElevenLabs currently leads the TTS space, but Reddit's LocalLLaMA community has been actively discussing the quest for "the best TTS solution that runs locally," with moss-nano and Kokoro among the candidates for edge devices.

The 2B-parameter scale appears aimed at balancing local runnability with quality, signaling the continued evolution of open TTS foundation models.

Casio's "Moflin" to Become a New Business Pillar as an AI Pet Robot

Casio Computer has revealed plans to grow its AI pet robot "Moflin" into a fourth business pillar, following the product's strong reception.

The move is intended to diversify operations alongside its consistently strong watch business, including the G-SHOCK line. With an elite organization reporting directly to the president charged with creating new businesses, Casio is aiming for a full-scale entry into the consumer AI robot market. The move is drawing attention as Casio explores a transformation from watchmaker into an AI robotics company.

Himeji City's Emergency Hospital AI Chatbot Achieves 90% Accuracy in Two Weeks

ITmedia reports that in a proof-of-concept trial conducted by Medialink with Himeji City, an AI chatbot for emergency hospitals achieved an accuracy rate of up to 90% in just two weeks.

While generative AI has been streamlining operations elsewhere, phone-based responses had long remained a scene unchanged for decades. The key to the trial was a "knowledge circulation" mechanism — a cycle that feeds back how AI responses were actually used in the field to continuously improve accuracy.

As a concrete set of lessons for making AI work in practice, the case offers valuable insights for other municipalities and healthcare institutions.

FT Warns of the Coming of "Anti-AI Populism"

In an article titled "The coming rise of anti-AI populism," the Financial Times points out that social backlash against AI could emerge as a new form of populism.

Against a backdrop of job losses to AI, impacts on the creative industries, and data privacy concerns, "anti-AI" voices could grow stronger across political divides. Discussion has also begun on Hacker News, and with the benefits of AI concentrating in a handful of companies, how society's overall acceptance evolves will be worth watching.

FP8 Quantization of Qwen3.6-27B Achieved on Dual RTX 3090s

Reddit's LocalLLaMA community reports that the "club-3090" project, targeting dual RTX 3090 setups, has experimentally added FP8 quantization support for Qwen3.6-27B.

The official Qwen/Qwen3.6-27B-FP8 model reportedly maintains performance nearly on par with the original BF16 version. The work demonstrates that FP8 quantization can make 27B-class models practically runnable even within the limited VRAM of a dual-3090 setup — a major step forward for the local LLM community.

Other Notable Moves

Two failure modes in multi-objective optimization of LLM judges**: Research published on Hugging Face provides a detailed analysis of two failure modes — "gradient dilution" and "instruction interference" — that arise when applying "textual gradients" for prompt optimization in multi-objective settings

Preseason.ai**: An open-source benchmark in which LLMs rank development tool options was announced on Hacker News

Running Gemma4 in Codex CLI**: Following Ollama's addition of Codex App support, experiments running local LLMs on Codex were reported on Qiita

From Vibe Coding to Vibe Business**: The concept of AI evolving from code generation to business operations is being discussed on Qiita

---