Midjourney Version 7's second update introduces a revamped editor interface, smart selection tool, and "Weird" parameter for creative image variations. While these features enhance artistic possibilities, they have sparked controversy over usability and workflow efficiency, with users divided on the balance between innovation and practicality.

SkyReels-V2, the world's first infinite-duration movie generation model based on a diffusion-forced framework, has been launched and open-sourced by Kunlun Wanwei. It addresses challenges in video generation such as motion quality, visual consistency, and length coordination through innovations like SkyCaptioner-V1, motion preference optimization, and efficient diffusion-forced techniques. The model excels in story generation, image-to-video synthesis, camera direction, and element-to-video task

Xunfei Spark X1 has been significantly upgraded, becoming the industry's only deep reasoning model trained on domestic computing power. It excels in tasks like math, coding, logic, and text generation, rivaling international standards despite fewer parameters. The model introduces a unified fast and slow thinking approach, enhancing customization and deployment simplicity. With advanced capabilities in language comprehension, problem-solving, and multimodal reasoning, Spark X1 supports developer

Microsoft released BitNet b1.58 2B4T, a 1-bit LLM with 2 billion parameters trained on 4 trillion tokens. It runs on commercial CPUs like Apple's M2 but requires Microsoft's bitnet.cpp framework and does not support GPUs. The model occupies only 400MB and outperforms other 1-bit models, aiming to make LLMs more accessible for edge devices and resource-constrained environments. Future plans include exploring larger models, increasing context length, supporting multiple languages, and co-designing

Microsoft Copilot Studio is introducing a new computer usage feature enabling AI agents to interact with third-party apps and websites via GUI, expanding automation possibilities beyond API connections. This user-friendly update leverages natural language processing, adapts to app changes automatically, and ensures security within Microsoft's infrastructure, allowing non-developers to automate tasks like data entry or invoice processing. Interested users can access the feature through an early w

OmiGPT is a budget-friendly, AI-powered wearable device driven by ChatGPT. Priced at $89, it features 64GB storage and connects via API to OpenAI's platform. Lightweight and versatile, it can be worn on the wrist or as a necklace. Developed by a San Francisco-based startup, OmiGPT prioritizes real-world utility over costly designs, aiming for continuous context-aware assistance through sensors and AI. Pre-orders are set to ship in Q2 2025.

WhatsApp is testing a new "Message Translation" feature in its Android beta version, enabling users to translate messages locally on their devices without sending data to external servers. The feature supports languages like Spanish, Arabic, and Hindi via downloadable language packs and allows manual or automatic translation. It maintains end-to-end encryption but may lack cloud-based accuracy. Feedback can be provided to improve translations, though no content is shared with Meta unless explici

OpenAI's study reveals o3 generates more hallucinatory content than o1 due to its dynamic strategy, raising accuracy concerns despite richer outputs. This highlights the need to balance creativity and reliability in future AI models.

Figma is developing an AI-powered app generator using Anthropic's Claude Sonnet model, enabling cross-modal input integration for creating interactive prototypes from text, design files, and images. This tool aims to lower app development barriers for non-technical users. Additionally, Figma is working on a new product called "Figma Sites" for website creation, signaling its expansion into full-stack digital product development with AI-driven solutions.

The Chatbot Arena, a crowdsourced AI model benchmarking platform, has rebranded as LMArena and launched a new company, Arena Intelligence Inc. Led by UC Berkeley researchers, it attracts one million monthly visitors for head-to-head model comparisons. The team aims to expand functionality, address usability issues, and maintain neutrality while exploring a business model focused on model evaluations. With improved features and experimental spaces, LMArena seeks to balance academic integrity and

TSMC reported Q1 revenue of NT$839.25 billion (US$25.8 billion), exceeding expectations despite lower mobile chip orders. High-performance computing drove 59% of sales, offsetting declines in smartphone and connected devices. Net income reached NT$361.56 billion (US$11.12 billion). Advanced processes (7nm+) accounted for 73% of wafer production, with 3nm chips at 22% of revenue. TSMC focuses on 2nm technology for future growth, set for mass production by year-end, offering improved energy effici

OpenAI introduces Flex processing to cut AI model costs by 50%, targeting non-production tasks. Compatible with o3 and o4-mini models, it is in beta with trade-offs like slower response times. Identity verification is now required for certain features and models to prevent misuse. This move counters rivals like Google, which released a cost-effective Gemini 2.5 Flash model.

Google offers free Google One AI Premium to US college students until June 30, 2026, with benefits including 2TB storage and AI tools like Gemini Advanced and Deep Research. Eligibility requires registration by June 30, 2025, using a valid .edu email. The package enhances academic productivity with features such as NotebookLM Plus, Veo 2, and Whisk, reflecting Google's strategic push in education amid growing AI competition.

Perplexity's AI assistant will feature prominently in Motorola's new Razr phone, set for announcement on April 24. It will coexist with Google's Gemini and have a dedicated interface. Perplexity is also working with T-Mobile's parent company on an "AI phone" and exploring potential integration with Samsung devices, though discussions are ongoing.

Wikipedia is addressing AI developers scraping content by collaborating with Kaggle to release a beta dataset optimized for training AI models. The dataset, available in English and French, includes structured content like abstracts, descriptions, and infobox data under an open license. It aims to reduce server strain from bots while benefiting small businesses and researchers. This initiative expands on prior content-sharing agreements with Google and the Internet Archive.

xAI enhances Grok chatbot with a new "memory" feature, allowing it to recall past conversation details for personalized responses. This beta feature is available on Grok.com and mobile apps, excluding EU/UK users. Unlike competitors like ChatGPT and Gemini, Grok offers transparent memory control, letting users view, delete, or disable saved data. xAI plans to integrate this functionality into the Grok experience on X.

Auradine Inc., a data center hardware startup, raised $153 million in Series C funding led by StepStone Group. The company specializes in energy-efficient servers for bitcoin mining and AI, featuring custom processors and cooling technologies like water-based systems. Their AH3880 system offers up to 600 trillion hashes per second with automatic performance adjustments. Auradine also launched AuraLinks AI, focusing on networking hardware for GPU clusters, aiming for five times the performance of

Windsurf, creator of a popular AI coding assistant, is in talks with OpenAI for a potential $3 billion acquisition. This could impact OpenAI's startup fund credibility, as it has invested in Cursor, another coding assistant provider. Recent signs include an email to Windsurf users about an upcoming announcement and a video from OpenAI's CPO highlighting Windsurf's capabilities. Founded in 2021, Windsurf has raised $243 million and generates $40 million in ARR, compared to Cursor's reported $200

OpenAI deploys a new monitoring system for o3 and o4-mini models to prevent providing harmful advice on biological and chemical threats. The safety-focused inference monitor, trained on content policies, identifies risky prompts and blocks responses 98.7% of the time in tests. Despite improvements, o3 and o4-mini pose higher risks than previous models in answering questions about developing biological weapons. OpenAI acknowledges limitations and continues to rely on human oversight while trackin

Anthropic's Claude AI now features autonomous multi-step research, Google Workspace integrations, and document ecosystem search. It can explore queries, cite sources, and provide context-aware responses using Gmail, Calendar, and Docs. Available in testing for select users, this update enhances enterprise functionality with security-focused document retrieval.