Today:

• OpenAI Expands Deep Research Access

• Microsoft Makes Copilot Voice and Think Deeper Free

• Google Launches Free Gemini Assist

• DeepSeek Fast-Tracks New AI Model

OpenAI Expands Deep Research Access

OpenAI is expanding access to its deep research tool, a web browsing agent that generates detailed reports, to all paying ChatGPT users. Subscribers to Plus, Team, Enterprise, and Edu plans get 10 queries per month, while Pro users now receive 120. This move intensifies competition with Google and Perplexity, which offer similar tools. Tech firms hope these features justify costly AI subscriptions while refining their persuasive capabilities.

Microsoft Makes Copilot Voice and Think Deeper Free

Microsoft is making Copilot Voice and Think Deeper, powered by OpenAI’s o1 reasoning model, free with unlimited use for all users. Previously restricted for free users, these features now allow extended AI conversations. Microsoft continues to offer its $20/month Copilot Pro for priority access and exclusive AI tools in Microsoft 365. Users may experience delays during peak times or security concerns.

Google Launches Free Gemini Assist

Google has launched a free version of Gemini Code Assist, offering 180,000 monthly code completions, far exceeding competitors like GitHub Copilot Free. Powered by Gemini 2.0, it supports all programming languages, integrates with major IDEs, and enhances code quality. This move intensifies the AI coding tool competition, making advanced AI-powered development tools more accessible to students, startups, and hobbyists without requiring a subscription.

DeepSeek Fast-Tracks New AI Model

Chinese AI startup DeepSeek is accelerating the release of its R2 AI model, aiming for an earlier launch before May. Following the success of R1, which disrupted global markets with its low-cost, high-performance AI, DeepSeek is focusing on improving coding and multilingual reasoning. The company, backed by its parent High-Flyer, has China’s strong support despite privacy concerns abroad. 

🧠RESEARCH

VideoGrain: Modulating Space-Time Attention for Multi-grained Video Editing

VideoGrain is a zero-shot video editing model that enhances fine-grained control using space-time attention modulation. It improves text-to-region alignment and reduces feature coupling, enabling precise edits at class, instance, and part levels. Experiments show state-of-the-art performance, making VideoGrain a breakthrough in multi-grained video editing.

DICEPTION: A Generalist Diffusion Model for Visual Perceptual Tasks

DICEPTION is a diffusion-based model that excels in multiple visual perception tasks using minimal data and computation. It achieves state-of-the-art results with just 0.06% of SAM-vit-h’s data. By unifying tasks through conditional image generation and color encoding, it enables efficient training and rapid adaptation with minimal fine-tuning.

Audio-FLAN: A Preliminary Release

Audio-FLAN is a large-scale instruction-tuning dataset designed to unify audio understanding and generation in large language models. Covering 80 tasks with over 100 million instances, it enhances zero-shot learning across speech, music, and sound. This dataset aims to bridge the gap in developing truly unified audio-language models.

CodeCriticBench: A Holistic Code Critique Benchmark for Large Language Models

CodeCriticBench is a comprehensive benchmark for evaluating the critique abilities of large language models in coding tasks. It assesses models on code generation and QA across varying difficulty levels, using fine-grained evaluation metrics. The benchmark addresses gaps in existing critiques, providing a more rigorous assessment of LLMs’ reasoning and feedback capabilities.

Thus Spake Long-Context Large Language Model

This survey explores the challenges and advancements in extending Large Language Models' (LLMs) context length to millions of tokens. It examines architecture, infrastructure, training, and evaluation, highlighting long-context LLMs’ potential and limitations. The paper also presents ten unresolved questions, offering a roadmap for future research in lifelong learning models.

🛠️TOP TOOLS

Niji Journey - AI-powered art generation application specialized in creating anime-style illustrations and artwork.

Knowbase AI - AI-powered tool designed to transform the way users store, access, and utilize their knowledge.

AskCodi - AI-powered platform designed to enhance the coding experience and provide comprehensive support for programmers.

BlackInk AI - Platform that has transformed the way tattoos are designed, leveraging the power of advanced artificial intelligence.

MyEssayWriter - Free AI Essay Writing tool

📲SOCIAL MEDIA

🗞️MORE NEWS

• Estonia is the first country to integrate ChatGPT Edu into its national education system, providing AI-powered tools to students and teachers. This initiative supports AI literacy, enhances learning, and prepares students for the future workforce.

• Chegg is suing Google, claiming its AI Overviews hurt traffic and revenue by keeping users on Google instead of visiting Chegg’s site. The lawsuit alleges unfair competition, monopoly abuse, and unjust enrichment, challenging Google’s AI-driven search model.

• Alibaba is releasing an open-source version of its AI model, Wan 2.1, for video and image generation. This move intensifies competition in China’s AI market as the company invests $52 billion in AI and cloud computing infrastructure.

• DeepSeek has resumed API access after server capacity shortages, allowing customers to top up credits. The Chinese AI startup, which claims to rival OpenAI at lower costs, sparked industry debates on AI affordability and accessibility.

• Over 1,000 musicians, including Kate Bush and Paul McCartney, released a “silent album” protesting proposed UK copyright changes. They fear AI companies will exploit their work without permission, threatening artists’ livelihoods and creative rights.