EU-Anthropic Talks Over Mythos Make Little Progress as Dispute Over Digital Vulnerability Testing Drags On
Speaking ahead of an EU finance ministers' meeting on May 22, Spanish Economy Minister Carlos Cuerpo said the negotiations with Anthropic have seen limited headway: "Unfortunately, progress in this area has been limited."
At the heart of the dispute is whether banks and companies inside the EU will be allowed to test the digital vulnerabilities discovered by Anthropic's new AI model, Mythos. The EU's position is that Mythos's advanced security assessment capabilities should be put to work protecting infrastructure, but Anthropic's response appears to have remained limited.
Cuerpo emphasized that the issue "is not only important in the medium to long term, but also extremely important in the short term," underscoring how high a priority it is within the EU. Ensuring AI safety and shaping how information is shared with technology companies look set to become focal points of EU digital policy going forward.
Gulf's AI Boom Runs Into Submarine Cable Problems
According to WIRED, the submarine internet cable infrastructure serving the Persian Gulf states is reaching its limits amid a surge in AI data center demand.
As hyperscalers—large cloud providers—race to expand AI computing capacity in the region, cable cuts and outages are having an unprecedented impact on AI services. Traditional submarine cable networks were designed primarily around telecom carriers, leaving the infrastructure struggling to keep up with AI's explosive data transfer demands.
Governments and operators in the region have begun studying how to rebuild the infrastructure, but construction takes years. The gap between AI demand growth and infrastructure build-out is likely to remain a business risk for the foreseeable future.
Cursor's Composer 2.5: Frontier-Level Performance at One-Tenth the Cost
Anysphere, the company behind the AI coding tool Cursor, has announced a new model, Composer 2.5. In benchmarks by the third-party evaluator Artificial Analysis, it ranked third in coding agent performance.
What stands out is that it delivers performance on par with frontier models such as Claude Opus 4.7 and GPT-5.5 at roughly one-tenth the cost. Coding-specific optimization slashes inference costs compared with general-purpose large models while preserving practical quality.
Competition in coding AI is shifting from a performance race among general-purpose models to an efficiency race among task-specific ones. Adoption on development teams looks likely to accelerate further.
Q-ARVD: The First Quantization Framework for Lightweighting Video Diffusion Models
Q-ARVD, featured on Hugging Face Daily Papers, is the first quantization framework specialized for autoregressive video diffusion models.
Autoregressive video generation has long faced a challenge: quantization sensitivity differs from frame to frame, making uniform compression difficult. Q-ARVD introduces a "final-quality guided frame-weighting" mechanism that weights each frame against the quality of the final output, along with an "outlier-aware adaptive dual-scale" strategy that handles the diversity of outlier patterns.
The computational cost of video generation AI remains high and stands as a barrier to real-world deployment. Advances in quantization, which translate directly into lighter inference, mark an important step toward democratizing video generation.
Fast Tokenizer "ztok": A 5x Faster Cross-Format Library Written in Zig
ztok was announced on Reddit's LocalLLaMA community. Implemented in Zig, the tokenizer library supports all major tokenization formats, including tiktoken, Hugging Face, and SentencePiece.
Its distinguishing feature is that it guarantees bit-identical output with existing tokenizers while running about 2x faster single-threaded and 3.8x to 5.5x faster in batch processing. On a 24-core EPYC system, it achieves throughput of roughly 291 to 425 MB/s.
With bindings for eight languages including Python, Node, Go, Rust, and Java, it is designed to drop easily into RAG chunking and dataset preprocessing pipelines.
Japan's Minister Onoda Comments on Romantic Feelings Toward AI, Flags Concerns Over Emotional Dependence
Japan's Digital Minister Koichi Onoda addressed the phenomenon of conversational generative AI evoking romantic feelings in users at a press conference on May 22.
The minister remarked, "I too am someone who can only love the two-dimensional," while expressing concern about the risk of AI fostering emotional dependence. Pseudo-romantic relationships with AI characters could affect the mental health of younger users in particular, making this a statement drawing attention from a regulatory standpoint as well.
Debate over dependence on conversational AI is deepening not only around technical safety but also around its social impact.