Controversy arises over Meta potentially using unpublished books from the LibGen database to train its AI models. An investigation revealed an upcoming book appeared in the database prematurely, raising concerns about intellectual property and data privacy. While platforms like NetGalley restrict access to such works, questions remain about Meta's compliance with laws regarding AI training data. This issue highlights the need for stronger protections of authors' rights.

OpenAI has temporarily limited image generation requests due to high demand for ChatGPT's image creation feature, straining GPU resources. CEO Sam Altman noted this is a temporary measure while improvements are made. Free users may soon generate up to three images daily. The GPT-4o model enhances realism and text rendering but requires significant computational power.

OpenAI will support Anthropic's Model Context Protocol (MCP), an open standard connecting AI assistants to data storage systems. This allows models to access various data sources for better responses. Several companies have adopted MCP, and OpenAI plans to integrate it into products like ChatGPT.

Nvidia is reportedly nearing a deal to acquire Lepton AI, a server rental company using Nvidia's AI chips, valued at several hundred million dollars. Lepton AI, founded two years ago, raised $11 million in seed funding in 2023. Meanwhile, competitor Together AI has raised over $500 million. Additionally, Nvidia acquired synthetic data startup Gretel last week.

OpenAI CEO Sam Altman announced the postponement of rolling out ChatGPT's image generation feature to free users due to its unexpected popularity. Introduced on Tuesday, the GPT-4o-powered tool allows in-app image creation, notably producing Hayao Miyazaki-style art. Currently, this feature is exclusive to paid subscribers with no set timeline for free user access.

OpenAI, the company behind ChatGPT, is projected to generate $12.7 billion in revenue this year, up from $3.7 billion last year, with revenue potentially reaching $29.4 billion next year driven by its paid AI software and a new $200/month subscription tier.

Kunlun Wanwei launched Mureka V6 and O1, advanced AI music models. Mureka V6 supports multi-language lyric writing and diverse music styles, while O1, the first CoT-integrated model, enhances coherence and artistic expression. Features include song referencing, voice cloning, and API services for developers. The MusiCoT algorithm boosts high-fidelity music generation, marking a significant industry advancement.

Amazon's Alexa Fund is expanding its focus to AI-driven technologies, investing in startups like Hedra, Ario, HeyBoss, and NinjaTech. These companies specialize in areas such as AI video generation, parental assistance, website design, and productivity tools. The fund has invested in over 100 companies, including Ring and Ecobee, emphasizing emerging tech and collaboration with innovative founders. Priorities include next-gen assistant experiences, embodied AI, and cost-efficient AI architecture

Microsoft introduced Researcher and Analyst in Microsoft 365 Copilot. Researcher integrates OpenAI's deep research model with advanced search capabilities for tasks like market strategy formulation. Analyst, based on o3-mini, excels in iterative data analysis and Python-based queries. These tools access work data and web resources, differentiating them from competitors. Challenges remain in preventing content hallucinations. Microsoft is offering early access via its Frontier program starting Ap

The Yuanbao Platform has updated its core models, releasing Hunyuan T1 and DeepSeek V3-0324. Hunyuan T1, Tencent's self-developed model, improves response speed and reasoning capabilities for processing lengthy content. DeepSeek V3-0324 excels in code generation and medium to long-form text creation. Users can choose between speed-focused options (Hunyuan Turbo S/DeepSeek-V3) or deep thinking models (Hunyuan T1/DeepSeek-R1). The platform aims to provide advanced AI tools.

Nvidia's G-Assist, initially a 2017 April Fools' joke, is now an AI assistant for RTX GPU users. Integrated into the Nvidia app, it optimizes gaming and system settings, supports voice/text commands, and controls lighting for compatible devices. It analyzes PC performance, suggests optimizations, and uses a local compact language model requiring 10GB storage. Compatible with RTX 30, 40, and 50 series GPUs (12GB VRAM minimum), G-Assist is an optional feature with new DLSS options.

Apple updated its policy on image collection for Apple Maps' Look Around feature, stating it will use data gathered from March 2025 onward to train models for image recognition and enhancement. Images captured via vehicles and backpacks will have faces and license plates blurred to protect privacy. These blurred images will be used for model training. Apple's AI-powered features, such as Image Playground and photo cleanup tools, may benefit from this initiative. Specific models to be trained are

OpenAI integrates new image generation into ChatGPT via GPT-4o, available across subscription tiers. It excels in attribute binding for up to 20 objects and improved text rendering, using an autoregressive method instead of diffusion. Practical applications include scientific diagrams, comics, and promotional materials. While slower than previous systems, the enhanced quality and embedded world knowledge are deemed worthwhile. Free tier limits align with DALL-E, though exact numbers are undisclo

Google updates AI models with Gemini 2.5, featuring improved base models and post-training techniques. Gemini 2.5 Pro, in beta, outperforms rivals in benchmarks and offers multimodal capabilities. It supports text, audio, images, video, and code, with plans for a 2 million token context window. The model's quality boost comes from a reasoning mechanism enabling step-by-step task processing for smarter responses.

DeepSeek V3-0324, the latest update of DeepSeek's V3 series, shows significant performance improvements despite being labeled a minor update. It excels in high-dimensional spatial problems and programming tasks, generating adaptive web pages comparable to Claude 3.7 Sonnet. Though not designed for reasoning, it demonstrates cognitive decomposition in complex scenarios. Remaining free and open-source under the MIT license, its 671-billion-parameter MoE model is available on HuggingFace and other

OpenAI updated ChatGPT's Advanced Voice Mode to make it more human-like and reduce interruptions during conversations. Free users can now pause without being interrupted, while paying subscribers enjoy enhanced AI personality and more direct, engaging responses. This improvement comes as competition in the AI voice assistant market grows, with new entrants like Sesame and Amazon's upcoming LLM-powered Alexa.

A new AI test, ARC-AGI-2, evaluates the general intelligence of AI models by presenting unfamiliar puzzles. Developed by the Arc Prize Foundation co-founded by François Chollet, it focuses on adaptability and efficiency rather than brute-force solving. Leading AI models score below 2%, while humans average 60%. The foundation's 2025 competition challenges developers to reach 85% accuracy on ARC-AGI-2 under $0.42 per task, pushing for more efficient and creative AI advancements.

Ant Group achieves AI cost reduction using domestic chips from Alibaba and Huawei, potentially rivaling NVIDIA's performance. Bloomberg reports these chips could threaten NVIDIA's market dominance, though demand for NVIDIA remains strong in China despite US restrictions. Neither party has commented officially.

OpenAI announces leadership changes: COO Brad Lightcap expands role, CEO Sam Altman focuses on tech strategy. Mark Chen promoted to Chief Research Officer, Julia Villagra to Chief HR Officer. The moves come amid recent departures and as OpenAI transitions to a for-profit entity, sparking legal action from co-founder Elon Musk.

French AI startup Mistral's CEO, Arthur Mensch, emphasizes the importance of nations developing their own AI infrastructure to avoid economic risks. He warns that countries without independent AI systems may face capital outflow, akin to nations lacking power plants a century ago. Mensch advocates for open-sourcing large language models to accelerate technological democratization, contrasting with closed-source approaches like OpenAI's. Mistral, valued at $6.2 billion, focuses on building fast g