🤖 AI Daily · Jul 16, 2026

Alibaba Qwen-Audio-3.0-Realtime tops speech reasoning; Thinking Machines unveils 975B Inkling multimodal model; OpenAI releases GPT-Red automated red-teaming with 84% success; xAI open-sources Grok Build coding agent; Apple Intelligence approved in China with Qwen integration; WPS Comate AI office client launched; Claude Code adds MCP connector calling; Ex-DeepMind researcher quits over military AI deal; OpenAI pushes 'reverse federalism' for AI safety; GPT-Red breaks nearly all prior models.

1. 1. Alibaba Qwen-Audio-3.0-Realtime Tops Speech Reasoning

   Alibaba's Tongyi Lab releases real-time speech model Qwen-Audio-3.0-Realtime, ranking #1 in Artificial Analysis Speech Reasoning, beating OpenAI.

2. 2. Thinking Machines Unveils 975B Inkling Multimodal Model

   Thinking Machines Lab launches Inkling, a 975B parameter multimodal model for text, image, and audio. 1M context window. Full weights to be released.

3. 3. OpenAI GPT-Red: 84% Attack Success, Crushes Human 13%

   OpenAI's GPT-Red uses self-play RL to simulate prompt injection attacks with 84% success vs human red team's 13%. Used to adversarially train GPT-5.6 Sol.

4. 4. xAI Open-Sources Grok Build Coding Agent

   xAI open-sources Grok Build, SpaceXAI's coding agent and TUI. Users can now compile and run fully locally pointing to local inference backends.

5. 5. Apple Intelligence Approved in China with Qwen Integration

   Apple Intelligence model licensed in China on Jul 8. Alibaba's Qwen will integrate as AI capability for Chinese iOS/iPadOS/macOS/visionOS users.

6. 6. WPS Comate AI Office Client Launched

   Kingsoft launches WPS Comate AI client with 6 modules: AI role experts, skill ecosystem, automated tasks. Supports both cloud and local modes.

7. 7. Claude Code Adds MCP Connector Calling

   Claude Code artifacts can now call MCP connectors, building dashboards and apps that fetch info and perform actions on demand. Available for Pro, Max, Team, Enterprise.

8. 8. Ex-DeepMind Researcher Quits Over Military AI Deal

   Researcher Alex Turner leaves DeepMind after Google signs unrestricted military AI deal with Homeland Security. His proposal to ban killer robots was ignored.

9. 9. OpenAI Pushes 'Reverse Federalism' for AI Safety

   OpenAI advocates 'reverse federalism' where states pass similar AI safety laws creating de facto national standards. CA, NY, IL have passed relevant bills.

10. 10. GPT-Red Breaks Nearly All Prior Models

    GPT-Red breaks nearly all prior models. Discovered vulnerabilities used to adversarially train GPT-5.6 Sol, boosting resistance from 0th to 90th percentile.