-
Google DeepMind made the Interactions API the default interface for Gemini models and agents. The older
generateContentAPI still works, but new agent features will ship through Interactions going forward. The interface now supports Managed Agents with Linux sandboxes, background execution, tool chaining with Google Search and Maps, and media generation for images, music, and speech. Developers can also choose Flex mode for lower cost or Priority mode for speed. -
Anthropic and Micron announced a partnership to co-design AI memory architecture. The companies plan to study how memory systems behave under different AI workloads and where they can improve performance and energy efficiency. Micron will supply HBM, DRAM, and SSDs, while also investing in Anthropic and deploying Claude internally. That structure also raises concerns about circular AI infrastructure deals, where investors become suppliers to the companies they fund.
-
Getty Images signed a multi-year licensing deal with OpenAI to bring licensed photos into ChatGPT search and discovery. The companies did not disclose financial terms or whether Getty content will be used to train future models. Getty’s path is notable: after pushing back against AI image generation, it launched its own generator, sued Stability AI, and is now leaning into licensed content for AI products.
-
MosaicLeaks shows how deep-research agents can leak private information through ordinary-looking external search queries. When agents combine private local documents with web retrieval, no single query may reveal a secret, but the accumulated query log can let an observer reconstruct sensitive facts. Prompting agents not to leak helped only inconsistently, and training only for task success made leakage worse. The proposed Privacy-Aware Deep Research method raised strict chain success from 48.7% to 58.7% while cutting answer/full-information leakage from 34.0% to 9.9%.
-
Hugging Face published a framework for benchmarking whether open models are “agentic enough” on real tooling. Instead of only checking the final answer, the benchmark looks at how much work the agent needed: turns, tokens, time, error recovery, and whether it used a clean path or a workaround. Using
transformersas the case study, the post argues that CLIs, Skills, and task-specific examples can make libraries much easier for coding agents to operate. -
PP-OCRv6 is now available on Hugging Face. PaddleOCR’s new general OCR family spans tiny, small, and medium models from 1.5M to 34.5M parameters, with small and medium supporting 50 languages. On PaddleOCR’s internal benchmarks, the medium model reaches 86.2% detection Hmean and 83.2% recognition accuracy. The focus is practical structured text extraction for document parsing, search, RAG, analytics, and agent workflows.
-
Microsoft is building a roughly 2-gigawatt data center campus in Pecos, Texas, with its own gas plant. Rather than waiting years for grid interconnection, Microsoft is funding on-site power for one of its largest capacity additions. The company says it will avoid raising local power prices and limit water impact, but the project highlights how AI infrastructure is becoming an energy and local-policy issue, not just a cloud expansion story.
Bottom line: today’s AI news is less about a single model release and more about the operational layer around AI: agent interfaces, memory and power infrastructure, licensed data, privacy-safe research agents, and tooling designed for autonomous systems.