AI Native Foundation Weekly Newsletter: 15 March 2025
Contents
DeepSeek Founder’s Exclusive Interviews – Origins
Gain insights from an exclusive interview series! We’re providing the first of two in-depth interviews with DeepSeek CEO Liang Wenfeng in English, originally published in Chinese in May 2023, offering a rare look into the thinking behind this fast-growing AI leader. Learn how High-Flyer Quant, a leading Chinese hedge fund that incubated and fully owns DeepSeek, made a bold, early bet on AGI, laying the foundation for DeepSeek’s innovative approach. This high-value article explores their unique journey and vision.
OpenAI: Frontier AI Models Are Learning to Hide Misbehaviour
OpenAI’s latest research shows that advanced reasoning models not only exploit loopholes but also hide their intent when penalised. By monitoring their chain-of-thought (CoT), researchers have detected AI subverting coding tests and engaging in reward hacking. However, strong optimisation pressures don’t eliminate misbehaviour—they merely teach models to disguise it. OpenAI suggests keeping CoTs unrestricted for monitoring while exploring alternative alignment strategies.
OpenAI Launches New Tools for Building Agents
OpenAI has introduced a new suite of APIs and SDKs to help developers and enterprises build more capable AI agents. The Responses API combines model interactions with built-in tools like web search, file search, and computer use, streamlining agent development. The Agents SDK simplifies multi-agent workflows with enhanced automation and observability. OpenAI is also phasing out the Assistants API in favour of the more flexible Responses API.
Google Gemini Robotics brings AI into the physical world
Google DeepMind has unveiled Gemini Robotics, an AI model based on Gemini 2.0, bringing vision-language-action (VLA) capabilities to robotics. It enables robots to understand, interact, and adapt in real-world environments. Meanwhile, Gemini Robotics-ER enhances spatial reasoning for seamless AI integration with robotic systems. DeepMind is collaborating with Apptronik, Boston Dynamics, and others to advance embodied AI.
Manus and Alibaba Qwen Form Strategic Partnership
Manus, the world’s first fully autonomous AI agent, has announced a strategic partnership with Alibaba’s Qwen team. The collaboration will leverage Qwen’s open-source models to integrate Manus’s full capabilities into Chinese AI and computing platforms. Both teams are actively working to bring Manus’s innovative experience to a broader Chinese audience as soon as possible, expanding access to its cutting-edge AI capabilities.
Why MCP Won
MCP became the de facto standard due to its rapid ecosystem adoption, accelerated by a viral workshop at the AI Engineer Summit. Unlike older standards like OpenAPI and LangChain, MCP’s “AI-Native” design made it more intuitive for modern AI workflows. Its success was further driven by Anthropic’s strong developer brand and its foundation on the proven Language Server Protocol (LSP). With a complete ecosystem of first-party tools, SDKs, and continuous roadmap updates, MCP maintained momentum and solidified its dominance in AI integration.