Нейродайджест за неделю (#131)

This week features significant updates in LLM capabilitiessingly assistantsethicsingd enoughto cover everything from new model releases like Gemini 3.8 up to advanced agentic workflows.

LLM

• Siri AI (Apple): Apple launched an English beta with personal context, screen understanding, and cross-app actions [Source].
• Gemini 3.8 Live (Google): Two voice models featuring visual context and background tool calls; Extended Thinking can reason while speaking [Source].
• Qwen3.8-Omni-Flash (Alibaba Cloud): A multimodal model accepting text, images, audio, and video with a 1M context window [Source].
• Jev (TypeSafe): A specialized model providing typed solutions instead of plain text or code within low latency [Source].

Генеративные модели

• Qwen-Image 2.1: A single 7B model for high-quality image generation and editing in 2K resolution [Source].
• Vidu S2: Enables real-time 720p avatar creation and live video stream editing such as character/background replacement [Source].
• Boreal (Creatify): An advertising focused text-and-image-to-video model trained on LTX-2.5 [Source].
• HiDream O1 Video: Offers way to create asynchronous enough videos using one reference promptor fixing duration at 10s [Source].
• Enhance Frame Rate (Runway): Toolset that increases frame rate up to 120 fps even though keeping original any qualityup_resolution [Source].

Агенты и продукты

• Claude Integration: Claude's chat, Cowork, and Design tools are being unified including editable Docs and Slides export features [Source].
• ChatGPT в Office: New add-in allows ChatGPT to work directly inside Word, Excel, and PowerPoint with context awareness [Source].
• Astra for Law (OpenAI): GPT-6 Astra has received legal search capabilities or specialized access restrictedly allowed certain firms under US law [Source].
• Агенты в DaVinci: Resolve Studio 21.1 now includes AI assistants integrated into the media library, editing, and rendering workflows [Source].

Безопасность и исследования

• Security Exploit/Hacktron: Researchers used Claude Opus to build an exploit targeting OpenAI via RCE and SSO errors [Source].
• Anthropic & Accenture Partnership:get special red teaming levelaccess ability through Faculty divisionto check safeguards specifically funded by Anthropic [Source].

! DYOR (Do Your Own Research)