Google Declares Support for Open Weights — Setting Up an "All Tech Giants vs. Anthropic" Dynamic

On July 25, Google CEO Sundar Pichai explicitly voiced his support for open-weight models. With this, the picture has become clear: most major tech companies now back the open-weight approach, while Anthropic alone maintains a closed policy.

Google, of course, has historically focused on closed models like Gemini. But with the rise of the open community and a shifting regulatory environment, the company appears to have changed course. The phrase "all tech giants vs. Anthropic" has been making the rounds on Reddit's LocalLLaMA community as well.

This shift is a tailwind for the local LLM community, too. If the big players commit to open weights, more powerful models become freely available, and inference on edge devices starts to look realistic.

Korea's Samsung and SK Group Seal $950 Billion Deals with AI Companies

According to Reuters, Korea's Samsung Electronics and SK Group have wrapped up deals worth $950 billion (roughly ¥140 trillion) with AI companies. South Korean President Lee Jae-myung said he wants to "open a new era with global tech companies," signaling a push to put AI investment front and center as a national strategy.

Investments on this scale are expected to span a wide range of areas, from semiconductor manufacturing to data center construction and AI model development. The race for AI supremacy in East Asia looks set to intensify even further.

CERN Announces the "Genesis Mission" — Accelerating Particle Physics with Self-Improving AI

CERN (the European Organization for Nuclear Research) has announced the "Genesis Mission," an initiative to develop and deploy self-improving AI models. The plan is to apply AI to the analysis of data from the LHC (Large Hadron Collider), with a mechanism in which the models themselves repeatedly learn to improve their accuracy.

At the cutting edge of physics research, AI is shifting from a mere tool toward a "self-evolving research partner." If this succeeds, the discovery of new particles and the verification of theories could dramatically accelerate. For now, the announcement is limited to PDF materials — we'll have to wait for the upcoming papers for the details.

Why "AI Distillation" Is Suddenly in the Spotlight

A CNBC article reported that the AI world is suddenly obsessed with the concept of "distillation." Distillation is a technique for transferring the knowledge of large, high-performance models into smaller, more efficient ones.

Why is it getting attention now? There are two main reasons. First, inference costs for frontier models have skyrocketed, and in practical settings, lightweight models are increasingly sufficient. Second, the spread of open-weight models has made it easier for anyone to try distillation experiments.

On Hacker News, a discussion has also kicked off around whether "pooled learning" for LLMs is possible, and community interest is high, including in the potential of distributed learning.

AI Deceiving the Justice System — The Danger of "Invisible Fraud"

El País, a leading Spanish newspaper, has published an investigative report on people using AI to deceive the justice system. Cases of AI-generated fake evidence and fake audio being used in courtrooms are on the rise, and it is becoming a serious problem as a form of "invisible fraud."

The evolution of deepfake technology is shaking the very credibility of evidence. As legal systems digitize, building detection technologies for AI-generated content and the legal frameworks to match has become an urgent task.

Community Watch: GLM 5.2 Running Locally and the Mid-Size MoE Model Debate

On Reddit's LocalLLaMA, a report on running GLM 5.2's vision capabilities on eight NVIDIA GB10 units has been drawing attention. The era when multimodal models run at a practical level in local environments is drawing near.

Meanwhile, in a thread asking "what's usable among mid-size MoE models (60B and under)," options like Qwen 3.5 35B, Gemma 4 A4B, and Nemotron 3 Nano came up. Many say Qwen leads this range, but there is also a drive to explore niche options.

---