AI Native Foundation Weekly Newsletter: 16 May 2025

GPT-4.1, optimized for coding and instruction following, is now available in ChatGPT. Pro, Plus, and Team users can access it via the ‘more models’ dropdown, while Enterprise and Education users will gain access in the coming weeks. The faster model serves as an alternative to o3 and o4-mini for coding tasks. Additionally, GPT-4.1 mini replaces GPT-4o mini for all users, with safety evaluation results available in the new Safety Evaluations Hub.

READ MORE

Alibaba’s QwenChat has officially released its Web Dev feature, enabling both technical and non-technical users to generate fully-functional, visually appealing websites using simple natural language commands. Users can create everything from e-commerce platforms to interactive applications without coding knowledge. The tool instantly produces complete, preview-ready website code from prompts as simple as “Write a fruit e-commerce website.” Available now at chat.qwen.ai under the WebDev option, this breakthrough makes web development accessible to everyone, though complex functionality may have limitations.

READ MORE

Google’s AlphaEvolve pairs Gemini models with automated evaluators in an evolutionary framework to discover optimized algorithms. The system quantifiably measures solution quality through objective metrics, enabling breakthroughs in mathematics and computing efficiency. Already deployed across Google’s ecosystem, it has improved data center scheduling, hardware design, and AI training performance. The technology has discovered novel matrix multiplication methods and solved geometric problems that have challenged mathematicians for centuries. Google plans an Early Access Program for academic researchers to further expand its capabilities beyond computing applications.

READ MORE

DeerFlow is ByteDance’s open source deep research framework built on LangStack (LangChain/LangGraph) with an optimized Multi-Agent architecture that significantly reduces token consumption and API calls compared to traditional Supervisor designs. Its Research Team mechanism supports multi-round dialogue, decision-making, and task execution with dynamic re-planning capabilities. DeerFlow also features MCP integration, AI-generated prompts via Meta Prompting, human-in-the-loop functionality, and unique output capabilities including podcast generation and PPT creation.

READ MORE

Meta FAIR is advancing machine intelligence with three major releases. OMol25, created with Berkeley Lab and other institutions, is the largest dataset for quantum chemistry calculations requiring 6 billion core hours. The Universal Model for Atoms (UMA) sets new standards for atomic interaction modeling across materials. Adjoint Sampling introduces reward-driven generative modeling without training data, while a collaboration with Rothschild Foundation Hospital reveals how neural language representations develop in children’s brains, showing striking parallels with large language models. These tools aim to revolutionize molecular design, healthcare, energy storage, and neuroscience research.

READ MORE

Transform meetings with Notion’s AI note-taking that automatically captures ideas, decisions and action items in 16 languages including English, French, and Japanese. Type /meet, join from Notion Calendar, or use automatic meeting detection to start. Your notes integrate seamlessly with projects and tasks at just $20/user—saving $138/user compared to equivalent standalone tools. Transcribes Zoom, Teams, and Google Meet with no bots required, while providing instant summaries, searchability, and privacy controls.

READ MORE

Tencent has open-sourced HunyuanCustom, a multimodal video generation tool that outperforms existing solutions in subject consistency. Built on HunyuanVideo, it enables users to generate videos by combining text, image, audio, and video inputs while maintaining high identity preservation. Key capabilities include single/multi-subject generation, audio-driven animation, and video editing. Unlike competitors that only animate images within original contexts, HunyuanCustom allows changing environments and actions while preserving identity. The tool is available on Tencent’s website with source code on GitHub.

READ MORE

Lovart is the world’s first design agent in beta that combines human creativity with AI on one canvas. Using a Talk-Tab-Tune workflow—speak your ideas, choose options, and refine designs—it transforms concepts into professional creations across various use cases from bag designs to marketing memes. This all-in-one platform eliminates tab switching and scales production by creating multiple deliverables from a single prompt. Join the waitlist to access this creative collaboration tool.

READ MORE