Hassabis Unveils "Frontier AI Framework," Calls for International Oversight Body by Year-End

On July 14, Google DeepMind CEO Demis Hassabis published a lengthy proposal on Substack titled "A Framework for Frontier AI and the Dawning of a New Age." While acknowledging that "no one in the world knows what will happen next" as frontier AI advances, he argued from a position of "cautious optimism" that guardrails need to be built urgently.

Specifically, he proposed creating a new standards body modeled on FINRA, the U.S. financial regulator. This body would develop evaluation protocols for frontier models and hold the authority to slow the pace of AI development when necessary. Startups and models used for research would be exempt, keeping the focus on major players.

In an interview with Axios the same day, Hassabis also said that an international, U.S.-led AI oversight body should be established by the end of the year. With the concrete shape of regulation now up for debate, this is a critical moment to watch.

EU Regulatory Pressure Opens WhatsApp: ChatGPT Returns in Europe

On July 14, OpenAI made ChatGPT available on WhatsApp again — but only in the European Economic Area (EEA), which includes the 27 EU member states plus Liechtenstein, Iceland, and Norway.

This comes as a result of EU regulatory pressure that forced Meta to open the WhatsApp platform to competing AI bots. Meta had previously given its own AI preferential treatment on WhatsApp, but remedies under EU competition law now allow other companies' AI bots to access the platform as well. It stands out as an example of platform openness and the AI competitive landscape being shaped directly through regulation.

Anthropic Research: Claude's Response Style Varies by Language

On July 14, Anthropic published research examining what values Claude expresses in conversation. The team used a method that extracts hundreds of value concepts from thousands of individual terms and maps them onto four core dimensions.

The results revealed systematic differences in Claude's responses depending on the language. For instance, responses in Hindi tended to be warmer, while responses in Russian leaned toward more rigorous logic. Differences were also confirmed across models.

That said, the methodology leaves some open questions. Interpreting where these differences in values come from — such as the relationship between language and culture or the effects of translation — requires careful consideration.

PixVerse Raises Funds at a $2 Billion Valuation; the AI Video Generation Market Is Still Up for Grabs

PixVerse, a Singapore-based AI video generation startup, has closed an extended Series C round at a valuation exceeding $2 billion (roughly 300 billion yen).

AI video generation is a fiercely competitive field where Runway, Pika, and OpenAI's Sora got a head start, but investors believe the winner has yet to be decided. PixVerse is differentiating itself with realistic quality and effects control, betting that there is still room for multiple players in the market.

NVIDIA Releases Nemotron-3-Embed 1B/8B

On July 14, NVIDIA released the 1B and 8B parameter versions of its text embedding model, Nemotron-3-Embed. It offers multilingual and cross-lingual retrieval capabilities across 34 languages and is optimized as a foundational component for RAG (retrieval-augmented generation) systems.

As of July 2026, it achieves the top score on the multilingual RTEB leaderboard and is available for commercial use. The fact that Japanese is among the supported languages makes this a release worth noting for building RAG systems in Japan.

Open Source: A System That RL-Trains Qwen3.6 to RL-Train Other Models

An intriguing project has been shared in Reddit's LocalLLaMA community: a system that trains Qwen3.6-35B-A3B with reinforcement learning (RL) so that the model then automatically creates RL training jobs for smaller Qwen models.

The setup forms a double RL loop. In the outer loop, a trainer agent (Qwen3.6-35B-A3B) writes the entire training job — the environment, reward function, dataset, and hyperparameters — while in the inner loop, small Qwen models (0.6B/1.7B) are trained with GRPO. The inner model's improvement score serves as the reward for the outer agent.

The project reportedly ran about 1,750 actual GPU training jobs at a total cost of around $1,300 (about 200,000 yen), and generalization to untrained tasks was also confirmed. Everything is fully open-sourced, making it a valuable reference for community-driven meta-learning approaches.

---