OpenAI, after raising $10 billion, spent most of it within 18 months and had to secure an additional $6.6 billion, planning to borrow $4 billion more. Annual spending is over $5.4 billion, projected to reach $37.5 billion by 2029, leading to a potential shift from non-profit to for-profit to attract more funding. AI development now relies on processing vast amounts of data, unlike traditional coding.

Infinigence AI introduces Megrez-3B-Omni, a 3 billion parameter local multimodal model for text, audio, and image analysis. It excels in benchmarks, supports multi-turn dialogues, and enhances privacy and efficiency on devices.

Writer launched Palmyra Creative, a large language model aimed at enhancing creative thinking and writing. It focuses on divergent thinking, collaborative creation, and targeted solutions, applicable in healthcare, finance, retail, and technology. Available through APIs, no-code tools, and Nvidia's NIM inference microservices.

Google updated Code Assist with Gemini 2.0, supporting longer context and larger codebases, and integrating with various data sources for real-time assistance within IDEs, enhancing developer productivity.

Nvidia launched NeMo Retriever, a multilingual microservice for AI engineers to build generative AI applications that store and retrieve data across languages, enhancing accuracy and context in multilingual environments.

OpenAI has no immediate plans to launch an API for its Sora video generation model, citing current capacity issues, despite competition from Google and AWS.

NVIDIA launched the Jetson Orin Nano Super Developer Kit, priced at $249, offering 67 TOPS AI processing power and 102GB/s memory bandwidth, suitable for AI and robotics projects.

OpenAI announced the o1 inference model, available to select developers, with enhanced features like function calling and image analysis. The o1 model is more expensive but offers self-fact-checking. New versions of GPT-4o and GPT-4o mini were also released, along with preference tuning and SDKs for Go and Java.

Meta AI and Stanford launch Apollo, a scalable video multimodal model supporting hour-long videos with up to 7B parameters. Innovations include scale consistency, efficient frame sampling, dual visual encoders, and ApolloBench. Apollo outperforms larger models in video understanding tasks, enabling advanced applications like video Q&A and content analysis.

Anthropic launched the open-source Model Context Protocol (MCP), enabling AI assistants to integrate with external data sources. The initiative has attracted developers and companies, fostering innovation despite facing deployment and security challenges.

YouTube introduces a feature allowing creators to authorize approved third parties like OpenAI and Microsoft to use their videos for AI training. Disabled by default, creators can enable it in YouTube Studio to support collaboration and AI development.

Apple accelerates foldable device development, including a foldable iPhone and a 2-in-1 laptop/iPad, targeting a 2026 release. Partnering with firms like Samsung, Apple aims to enhance design and address display challenges to boost sales.

MidJourney introduces personalized features like an Inspiration Board for curated uploads and multi-profile support. Enhanced image sorting speeds and streamlined profile creation enable tailored AI art generation, improving user workflows and collaboration.

DeepMind launches Veo 2, an AI that generates 4K videos over two minutes, surpassing OpenAI’s Sora in resolution and length. Currently in testing with VideoFX, Veo 2 offers enhanced realism and camera control but still faces coherence challenges. It will expand via Google’s Vertex AI and includes content filters and watermarking to prevent misuse.

Meta introduces three new features for Ray-Ban smart glasses: real-time AI and live translation for early access members, and Shazam music recognition available in the US and Canada. The AI assistant can suggest recipes and supports English, Spanish, French, and Italian. Users need software updates to access these features.

Google unveils Whisk, an AI tool that generates images from reference photos instead of text prompts, enabling rapid visual exploration and prompt refinement. Powered by Imagen 3. Additionally, Google launched Veo 2, a video AI model with enhanced capabilities, available via VideoFX and expanding to YouTube Shorts next year.

OpenAI announces ChatGPT search is now free for all users with an account. Mobile search enhanced with images, ratings, and maps. New feature prioritizes website links before details. Introduces Sora text-to-image model and a $200/month ChatGPT Pro subscription.

Elon Musk’s X users gain free access to xAI’s advanced Grok-2 chatbot, featuring increased speed and multilingual support. xAI’s valuation rises to $50B after funding, amid legal disputes with OpenAI. Concurrently, Google’s Project Moohan propels XR integration with leading headsets, fostering payment innovations across industries.

Researchers from Tsinghua and CMU developed optimal inference strategies for LLMs, demonstrating that smaller models can match larger ones using methods like REBASE. This advances computational efficiency and challenges traditional scaling notions in AI.

Meta unveiled Video Seal, an invisible neural watermarking system to protect AI-generated videos. Open-sourced under the MIT License, it includes tools for watermark comparison and application, enhancing copyright protection.