The future of presentations, powered by AI

Gamma is a modern alternative to slides, powered by AI. Create beautiful and engaging presentations in minutes. Try it free today.

Today:

• White House Tightens Export Limits

• Microsoft Unveils New AI Division

• OpenAI Proposes US AI Strategy

• Google Cloud Launches Automotive AI

• AI Startup Blaize Goes Public

• Mistral Unveils Faster Code Model

White House Tightens Export Limits

The White House has introduced strict regulations limiting the export of Nvidia's advanced AI chips, aimed at controlling computing power sold to most countries. The rules, effective in a year, allow exemptions for businesses adhering to specific security and human rights standards. These measures face resistance from the industry and will ultimately be implemented under the Trump administration. The move highlights ongoing tensions over AI dominance and trade control.

Microsoft Unveils New AI Division

Microsoft has launched CoreAI — Platform and Tools, a new AI division led by former Facebook engineering head Jay Parikh. This division aims to develop foundational AI technologies and tools for Microsoft and its customers, reflecting a transformative shift in AI. CEO Satya Nadella emphasized AI's role in reshaping applications. The company continues its strategic AI investments, including its partnership with OpenAI, while competing with Amazon, Google, and others in the growing AI landscape.

OpenAI Proposes US AI Strategy

OpenAI’s Economic Blueprint outlines strategies for the U.S. to lead AI innovation while ensuring equitable access, safeguarding national security, and driving economic growth. Highlighting the transformative potential of AI, it emphasizes the need for common-sense, nationwide regulations to foster innovation and prevent misuse. The plan advocates for investments in chips, data, energy, and talent to counter global competition and maximize AI’s benefits for all communities.

Google Cloud Launches Automotive AI

Google Cloud has launched the Automotive AI Agent, designed for automakers to create advanced, generative AI-powered in-car assistants. Built with Gemini and Vertex AI, it supports natural, multilingual conversations and handles complex queries. Mercedes-Benz will debut the agent in its MBUX Virtual Assistant for the upcoming CLA, enabling personalized navigation, real-time traffic insights, and multi-turn dialogue. The innovation enhances user experience with intuitive, conversational capabilities tailored to driver and passenger needs.

AI Startup Blaize Goes Public

AI chip startup Blaize, founded by ex-Intel engineers, will go public on NASDAQ via a SPAC merger, valuing the company at $1.2 billion. Specializing in edge AI chips for smart devices, Blaize focuses on low-latency, power-efficient solutions for physical-world applications like drones and security cameras. Despite recent financial losses, the company has $400 million in pipeline deals. This marks a shift from data-center-centric AI to integrated, real-world use cases.

Mistral Unveils Faster Code Model

French AI startup Mistral has released Codestral 25.01, a faster and more versatile version of its code generation model. Supporting over 80 programming languages, it doubles the speed of its predecessor and excels in benchmarks for fill-in-the-middle tasks, ranking second in LMsys' Copilot Arena. Accessible via IDEs like VS Code and JetBrains, or APIs from Mistral, Google, Microsoft, and Amazon, Codestral offers enterprise deployment and a free trial for developers.

🧠RESEARCH

VideoRAG: Retrieval-Augmented Generation over Video Corpus

VideoRAG is a framework that enhances Retrieval-Augmented Generation (RAG) by dynamically retrieving videos based on query relevance. Unlike prior approaches, it leverages both visual and textual video content for richer, more accurate outputs. Using Large Video Language Models, VideoRAG outperforms existing methods in utilizing videos as a multimodal knowledge source.

OmniManip: Towards General Robotic Manipulation via Object-Centric Interaction Primitives as Spatial Constraints

OmniManip introduces an object-centric framework for robotic manipulation in unstructured environments, bridging Vision-Language Models (VLM) reasoning with precise 3D spatial tasks. By leveraging object affordances, it translates VLM reasoning into actionable constraints without fine-tuning. A dual-loop system ensures robust, real-time control and demonstrates strong zero-shot generalization in diverse manipulation tasks.

LlamaV-o1: Rethinking Step-by-step Visual Reasoning in LLMs

LlamaV-o1 enhances visual reasoning in large language models (LMMs) with a step-by-step approach. It introduces a new benchmark for multi-step visual reasoning, a detailed step-focused metric, and a multimodal model trained via curriculum learning. LlamaV-o1 outperforms competitors in accuracy and speed, advancing interpretable, sequential problem-solving capabilities in visual contexts.

Enabling Scalable Oversight via Self-Evolving Critic

SCRIT (Self-evolving CRITic) is a framework enabling LLMs to self-improve critique capabilities without external oversight. By generating synthetic data, employing contrastive self-critique, and validating quality through correction outcomes, SCRIT enhances critique and error identification. Tested on Qwen2.5-72B-Instruct, it achieves a 10.3% performance boost, demonstrating scalability and superiority over existing methods.

ConceptMaster: Multi-Concept Video Customization on Diffusion Transformer Models Without Test-Time Tuning

ConceptMaster advances Multi-Concept Video Customization (MCVC) by addressing identity mixing and data scarcity issues in text-to-video diffusion models. It introduces decoupled multi-concept embeddings for maintaining concept fidelity and a systematic pipeline for collecting diverse, high-quality data. ConceptMaster outperforms existing methods, excelling in fidelity, identity decoupling, and video generation across various scenarios.

🛠️TOP TOOLS

ChefGPT - AI-powered culinary assistant designed to revolutionize home cooking and meal planning.

PhotoRoom - AI-powered photo editing app that specializes in quick and easy background removal and image enhancement.

Aomni - AI-powered B2B sales intelligence platform designed to streamline account research and enhance sales efficiency. 

Magic ToDo - AI-powered task management tool designed to simplify complex tasks for neurodivergent individuals and anyone seeking enhanced productivity. 

QuillBot AI - AI-powered writing assistant designed to enhance and streamline the writing process.

📲SOCIAL MEDIA

🗞️MORE NEWS

• Alibaba introduced Qwen Chat, a free web interface for testing Qwen AI models, offering tools for conversation, programming, reasoning, and image analysis. Future plans include web search, image generation, and voice interaction features.

• NVIDIA and IQVIA are partnering to create healthcare-focused AI agents using NVIDIA's AI Foundry platform. These tools aim to streamline clinical trials, improve drug development efficiency, and enhance patient outcomes while prioritizing compliance and privacy.

• AWS and General Catalyst announced a transformative partnership to revolutionize healthcare using generative AI. Their collaboration focuses on personalized care, efficiency, diagnostics, and real-world clinical applications, leveraging AWS’s AI technologies and General Catalyst’s health expertise.

• The UK government unveils a £14 billion AI action plan, featuring data center growth zones, a new supercomputer, and private investments from Vantage, Nscale, and Kyndryl, aiming to bolster AI infrastructure, jobs, and innovation.

• NVIDIA criticizes the Biden Administration’s proposed “AI Diffusion” rule, arguing it threatens U.S. technological leadership by imposing restrictive regulations on computing and AI innovation, undermining global competitiveness and economic growth.