QvQ, a 72B-parameter open-model by Qwen, excels in multimodal reasoning with enhanced accuracy and efficiency, showing strong performance on benchmarks like Visual7W and VQA. Its flexible and scalable design supports broad applications and minimal fine-tuning.

Google compared Gemini AI's responses with Anthropic's Claude during evaluation, raising questions about authorization. Contract workers scored each model on accuracy and informativeness, noting Claude's emphasis on safety. Google stated it did not use Claude to train Gemini.

At CES 2025, Samsung will highlight AI-integrated home appliances, particularly smart refrigerators with advanced food management and user interaction. The focus is on AI, with less emphasis on mobile devices.

OpenAI is reportedly reconsidering developing its own humanoid robot, following the quiet shutdown of its robotics division in 2021. The company has invested in related startups like Figure and 1X, and could leverage these partnerships to catch up in the competitive field.

OREO, an offline RL method, enhances LLMs' multi-step reasoning by optimizing the soft Bellman equation, training policy and value models simultaneously, and improving credit assignment. It outperforms existing methods on benchmarks like GSM8K, MATH, and ALFWorld, showing significant accuracy improvements and better generalization.

Coralogix acquires Aporia to enhance AI system monitoring and security, integrating Aporia's tech into its services and launching a dedicated AI research center.

Microsoft is reducing its reliance on OpenAI for Microsoft 365 Copilot, planning to integrate custom and open-source AI models to lower costs and improve response times. Phi-4, with 14 billion parameters, and Meta's Llama series are potential candidates.

Nvidia, with a market value over $3.4 trillion, will highlight the RTX 5000 series GPUs at CES 2025. CEO Jensen Huang's keynote on Jan 6th will cover AI, robotics, and automotive applications, attracting global tech attention.

xAI, founded by Elon Musk, raised $6 billion in a funding round with 97 investors, bringing total funding to $12 billion. The company aims for a $50 billion valuation and is integrating its AI model Grok into the X platform, adding features like chatbots and image generation. xAI faces competition from OpenAI and Anthropic and plans to expand its data center and GPU capacity.

The o3 AI model has achieved impressive results in the ARC-AGI benchmark but at a high computational cost, making it impractical for everyday use. While groundbreaking, its economic viability is questioned, and efforts are underway to develop more efficient AI inference chips.

xAI is testing an independent iOS app for its Grok chatbot, now available in Australia and a few other countries. The app, which previously was exclusive to xAI users, offers web access, real-time data, and generative AI features like text rewriting, summarization, Q&A, and image generation. An Android version and a dedicated website at Grok.com are also in the works.

Google's Deep Research tool now supports 45 languages and is available to Gemini Advanced subscribers in over 150 countries, offering comprehensive research capabilities and source links for efficient and accurate information gathering.

ASUS unveiled the NUC 14 Pro AI, a mini PC with Intel Core Ultra 9, up to 32GB memory, and Copilot Plus support. It offers 5 CPU configurations, impressive GPU and NPU performance, and is more compact than Mac Mini. Release date and price are not yet announced.

Google's Files app now lets users activate Gemini to ask questions about PDFs, requiring a Gemini Advanced subscription. This feature, introduced at Google I/O 2023, is being gradually rolled out.

OpenAI previewed its latest advanced reasoning models, o3 and o3-mini, at the "12 Days of OpenAI" event. These models excel in coding, mathematics, and science, with o3 outperforming predecessors in benchmarks. OpenAI also introduced deliberative alignment for enhanced safety, requiring models to reason through safety decisions. The company is accepting applications for testing, with a cautious release plan.

ExploreToM, a framework using the A* search algorithm, generates diverse and challenging datasets to evaluate and enhance Theory of Mind (ToM) in AI. It reveals current LLM limitations and improves performance through fine-tuning, supporting complex social reasoning.

Patronus AI launched Glider, a 3.8 billion parameter open-source model for evaluating large language models, outperforming larger tools in accuracy and providing transparent, real-time assessments with high-quality reasoning.

Microsoft is previewing a real-time translation feature on Copilot Plus for Intel, AMD, and Qualcomm platforms, supporting over 44 languages into English and some into Simplified Chinese.

Kuaishou's Keling 1.6 model enhances text responsiveness, visual aesthetics, and motion coherence, offering standard and high-quality modes. High-quality mode improves video quality by 195% with 1080p resolution, while standard mode provides 720p. Non-members get monthly inspiration points, and more are available with a membership.

Instagram plans to launch an AI-powered video editing feature next year, allowing users to modify videos with text commands. Based on Meta's Movie Gen AI, it aims to provide creators with more tools without needing professional skills. The feature was previewed by Instagram head Adam Mosseri, showing changes in attire, background, and appearance. Despite a smooth demo, the true performance is uncertain until the official release.