This week saw a mix of breakthroughs and security concerns in AI. OpenAI paused Pro subscriptions due to infrastructure strain from its Astra model, while Anthropic reported large-scale distillation attacks from Chinese labs. Meanwhile, Google DeepMind released an atlas of all possible DNA variants, and IBM open-sourced a time series model. Safety and efficiency were recurring themes, from boundary-aware refusal to agentic video analysis that cuts token use.

  1. OpenAI pauses Pro subscriptions as Astra demand strains systems - OpenAI temporarily halted new sign-ups for its $200 monthly Pro plan because demand for the Astra model overwhelmed infrastructure. This shows that even leading AI providers face capacity limits when releasing highly capable models, and it may push users to alternative platforms or lower tiers.
  2. Anthropic details distillation attacks from Chinese AI labs - Anthropic reported nearly 200 million exchanges linked to distillation campaigns from Alibaba, Moonshot AI, and DeepSeek targeting Claude's reasoning. This highlights the ongoing risk of model theft and the need for stronger protections against unauthorized training on proprietary outputs.
  3. AlphaGenome atlas maps all 9 billion DNA letter changes - Google DeepMind released a free atlas predicting molecular effects of every possible single-letter DNA variant in the human genome. This resource could accelerate genetic research and personalized medicine by providing a comprehensive reference for variant interpretation.
  4. IBM releases Granite time series model with permissive license - IBM's Granite Time Series PatchTST-FM-R2 offers strong zero-shot forecasting under Apache 2.0 and OpenMDW licenses. The permissive licensing makes it accessible for commercial use, potentially lowering barriers for businesses needing time series predictions without training custom models.
  5. Gemini adds agentic video analysis to cut token use - Google DeepMind launched agentic video understanding for Gemini Flash models, reducing token consumption by up to 88% and costs by up to 66% while improving accuracy. This makes video analysis more affordable and scalable for applications like surveillance, content moderation, and media indexing.
  6. Safety tuning needs boundary-aware refusal, not topic bans - New research shows that safety tuning should refuse only harmful subsets of a topic, not the whole topic, to avoid over-refusal on safe prompts. This approach could improve user experience by reducing false refusals while maintaining safety, a key challenge for deployed models.

The strongest shared signal this week is the tension between capability and control. As models become more powerful and widely used, providers face infrastructure limits, security threats, and the need for more nuanced safety mechanisms. Efficiency gains, like those in video analysis, and open resources, like AlphaGenome and Granite, point toward broader access, but the industry must also address misuse and over-refusal to build trust.