AI Daily — June 19, 2026
2026-06-19
rss.arxiv.org
DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
DeepSeek releases a preview of two new MoE models: DeepSeek-V4-Pro (1.6T total / 49B activated parameters) and DeepSeek-V4-Flash (284B total / 13B activated), both supporting 1M-token contexts. Key architectural innovations include Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) for long-context efficiency, Manifold-Constrained Hyper-Connections replacing standard residual connections, and the Muon optimizer for training stability. This represents a significant scaling step for open-weight MoE models with practical long-context capabilities.
openai.com
Using AI to help physicians diagnose rare genetic diseases affecting children
Researchers used an OpenAI reasoning model to assist in diagnosing rare pediatric genetic diseases, yielding 18 new diagnoses in previously unsolved cases. The work demonstrates reasoning models' practical utility in navigating the large, sparse evidence space typical of rare disease genomics. This is a concrete clinical validation of large reasoning models in a high-stakes diagnostic setting.
openai.com
Improving health intelligence in ChatGPT
OpenAI details how GPT-5.5 Instant improves health and wellness responses in ChatGPT, citing stronger multi-step reasoning, better context retention, and clearer lay-language communication. Evaluations were conducted with physician input to assess accuracy and safety of health-domain outputs. The update signals OpenAI's continued investment in making general-purpose models more reliable in regulated, high-stakes domains.
huggingface.co
MosaicLeaks: Can your research agent keep a secret?
ServiceNow researchers introduce MosaicLeaks, a benchmark probing whether agentic research systems inadvertently leak sensitive or proprietary information through mosaic-style inference across multiple retrieved documents. The work highlights a novel security surface in RAG and multi-step research agents that goes beyond standard data leakage concerns. This is relevant for enterprise deployments where agents traverse internal knowledge bases.
www.anthropic.com
Anthropic opens Seoul office and announces new partnerships across the Korean AI ecosystem
Anthropic has established a Seoul office and announced partnerships with Korean technology and enterprise partners, expanding its footprint in the Asia-Pacific region. The move follows similar regional expansions by OpenAI and Google and reflects growing enterprise AI demand in South Korea. No new model or technical releases were announced alongside this geographic expansion.