Google's Gemini Live screen sharing feature, previously a paid option for Pixel 9 and Galaxy S25 users, is now free for all Android users. Leveraging AI to interact with camera and screen content, the feature was initially planned as a paid subscription. Positive feedback prompted Google to expand access. Microsoft's Copilot Vision is also now free in Edge.

Microsoft Copilot's Vision feature is now free in Edge, enabling voice interaction to assist with on-screen tasks. While system-level Copilot Vision remains Pro-exclusive, users can activate the free version via a link, granting real-time screen analysis for recipes, interviews, or gaming. Activation may require multiple attempts, and interface issues might occur depending on device performance. Microsoft assures no user data is collected during sessions.

OpenAI launched two new reasoning models, o3 and o4-mini, featuring powerful logic and efficient performance. These models introduce image reasoning capabilities, enabling visual problem-solving by integrating and manipulating images. They are also integrated with ChatGPT tools, supporting web browsing and image generation. The release expands OpenAI's AI capability matrix following the GPT-4.1 announcement. Older models like o1 and o3-mini will phase out gradually.

Microsoft's Copilot Studio introduces "Computer Usage," a new feature enabling AI agents to interact with websites and desktop apps by clicking, selecting, and entering text. Similar to OpenAI's Operator, it automates tasks without needing API connections, supporting data entry, market research, and invoice processing. The tool adapts to UI changes seamlessly and offers more flexibility than the consumer-focused "Actions" feature.

Apple is doubling down on mixed reality with plans for two new Vision Pro devices while keeping its long-term focus on stylish AR glasses. One headset will be a lighter, more affordable version for everyday use, while the other will be a wired model for ultra-low latency tasks. Despite the high price and bulkiness of the current Vision Pro, CEO Tim Cook remains committed to developing practical AR glasses that integrate seamlessly into daily life, viewing it as a top priority for the company's f

OpenAI is exploring a social networking platform with a prototype featuring ChatGPT's image generation tool. While uncertain about entering the social media space, the project could capitalize on ChatGPT's large user base and create new revenue streams via advertising. With Meta facing an antitrust trial, OpenAI might position its platform as an Instagram alternative. CEO Sam Altman is reportedly inviting external testers for evaluation, but details on integration and release remain unclear.

OpenAI acquires Context.ai, a startup specializing in AI model evaluation. Co-founders Henry Scott-Green and Alex Gamble will join OpenAI to develop model assessment tools. This move underscores the growing importance of robust evaluation metrics as AI models become more complex. Context.ai's tools analyze model interactions, helping developers improve performance and reliability. Financial details are undisclosed, and it's unclear if all Context.ai employees will join OpenAI. The acquisition hi

Grok, developed by xAI, launched Grok Studio, a document editing and creation tool accessible to all Grok.com users. It supports collaboration on documents, code, reports, and games with features like HTML preview and code execution in Python, C++, and JavaScript. A recent update also added Google Drive integration for file attachments.

OpenAI launches a new non-profit advisory board with members Dolores Huerta, Monica Lozano, Dr. Robert K. Ross, and Jack Oliver to guide its charitable initiatives. This move follows OpenAI's transition to a for-profit entity, which faced criticism from former employees and industry leaders. The advisory board aims to expand influence and ensure the non-profit division supports global issues like healthcare and education.

Cohere Inc. launched Embed 4, an AI model for search and retrieval in assistant-based systems. It transforms documents into vectors, supporting over 100 languages and handling multimodal data like text, images, and graphs. With a context length of 128,000 tokens, it processes long documents and noisy data, benefiting industries like finance and healthcare. Agora, a Cohere customer, uses Embed 4 to improve e-commerce search. The model is integrated into Cohere's North platform and available on Az

OpenAI adds an image library feature to ChatGPT, allowing all users to access and manage AI-generated images easily. Available on mobile and web, it displays images in a grid format with a button to create new ones, enhancing management for artistic or daily use.

Anthropic is set to launch "Voice Mode," a voice feature for its Claude AI chatbot, competing with OpenAI's ChatGPT. The update may roll out this month with three English voice options: Airy, Mellow, and Buttery. Previously confirmed internally by Anthropic’s CPO, Mike Krieger, the feature was discovered in the iOS app by researcher M1Astra. Bloomberg verified the findings, though Anthropic has not yet commented. This move comes as Anthropic, founded by ex-OpenAI employees, expands its offerings

Google is offering trial access to its AI video generator, Veo 2, for Gemini Advanced subscribers. This text-to-video model creates 8-second, 720p clips with cinematic realism and embedded digital watermarking. Subscribers can generate videos via prompts on web and mobile, with a monthly limit. The upgraded model improves physics understanding and animation quality. Additionally, Google One AI Premium users can access Whisk Animate, converting images into Veo 2 videos globally through Google Lab

Anthropic launched two key features: Claude's integration with Google Workspace and a new research tool. The AI assistant can now summarize emails, identify tasks, and find relevant documents in Gmail, Calendar, and Docs, competing with Microsoft Copilot. Its agent-based research capability performs interconnected searches, providing cited results within 1-5 minutes to enhance problem-solving without workflow disruption.

GPT-4.5 passed the Turing Test by convincing 73% of participants it was human in text-based conversations, according to a UC San Diego study. It outperformed predecessors like GPT-4.0 and other models. Researchers warn of potential misuse, such as misinformation or fraud. OpenAI plans to replace GPT-4.5 with GPT-4.1 this summer. The Turing Test remains relevant for assessing machines' ability to mimic human-like interactions.

Extensive pre-training of language models can lead to "catastrophic overtraining," where increased training tokens reduce model performance. Researchers from CMU, Stanford, Harvard, and Princeton studied OLMo-1B with 2.3 trillion vs. 3 trillion tokens, finding a 3% performance drop in the latter. Beyond an inflection point, models become fragile, reversing gains and making fine-tuning harder. Adding Gaussian noise confirmed similar declines. Developers should determine optimal training limits or

Google unveiled DolphinGemma, an open-source AI model collaborating with Georgia Tech and the Wild Dolphin Project to decode dolphin communication. Trained on decades of data, it analyzes clicks, whistles, and pulses, generating sequences resembling dolphin sounds. Integrated with the CHAT system, it aims to establish a shared vocabulary using synthetic sounds. With around 400 million parameters, the model runs on smartphones for real-time analysis, reducing reliance on custom hardware. This bre

Hugging Face acquires Pollen Robotics, marking its first major move into hardware sales. The French startup behind the open-source humanoid robot Reachy 2 joins Hugging Face, adding 30 employees. This aligns with Hugging Face's growing focus on robotics, aiming to reduce costs and potentially open-source hardware designs. Co-founder Thomas Wolf emphasizes an open-source approach for safer, interactive robotics, expanding beyond AI models and datasets.

OpenAI may soon update its model naming convention, aiming for a more refined approach by summer. CEO Sam Altman acknowledged public jokes about current names like "GPT-4o," indicating a few months left before the new system is finalized. The revised strategy seeks to enhance how OpenAI's technological advancements and product identity are perceived, though public reception remains uncertain.

Meta plans to use EU user data from Facebook and Instagram to train AI systems, focusing on public posts, comments, and chat records while excluding private messages. The initiative aims to improve regional adaptability for multimodal AI. Meta will notify users via in-app alerts and emails, offering an opt-out option. Implementation remains delayed pending regulatory feedback, following last year's suspension of AI training efforts in Europe due to Irish authority concerns. Similar training usin