The Holidays come to OpenAI, while other foundation companies launch updates
what to know for now
🎄 OpenAI kicks off holiday AI festivities. OpenAI's "12 Days of OpenAI" event is delivering daily updates, starting with the o1 reasoning model, ChatGPT Pro, and hints of new features like video input for Advanced Voice Mode. Sora, a groundbreaking text-to-video AI generator, is also anticipated this week alongside incremental updates across their offerings. Read more
📉 Smaller, cheaper, more powerful Llama 3.3. Meta's Llama 3.3 delivers advanced performance on multilingual tasks with 70B parameters, rivaling larger models like Llama 3.1-405B while requiring significantly less GPU memory, reducing costs by up to $600,000 in hardware. This open-source model emphasizes cost-effective inference, sustainability, and safety features, supporting a wide range of applications with a 128k token context window and advanced alignment techniques. Read more
📢 AWS talks through advancements at re:Invent. Amazon had several announcements at AWS re:Invent 2024, including the Amazon Nova foundation models, Trainium3 chips, and enhancements to Amazon Bedrock and SageMaker. New offerings streamline Gen AI development, boost efficiency, and expand capabilities across data centers and databases. Read more
🪙 Gemini-exp-1121 excels with 2M tokens. Google introduced its Gemini-exp-1121 model, boasting a 2 million token context window, now previewing on AI Studio and ranking high on LM Arena. This release precedes the anticipated Gemini 2.0, set to bring further advancements. Read more
🧪 AI Research of the Week
Boundless Socratic Learning with Language Games
From Google DeepMindJake’s Take: The paper proposes using Socratic learning (an educational method that uses questions to find answers) for AI, in an effort to improve themselves within closed environments using language as both input and output. It highlights three key needs: clear feedback, broad data coverage, and enough computational power. The authors propose "language games" as a way to ensure continuous learning and alignment without external inputs.
This type of framework can help works towards AGI (or even ASI) by reducing dependency on external data sources. However, it would also make challenges around responsible AI alignment and system robustness that much harder.
The industry's inclination to chase scale over caution could make implementing Socratic learning both an exciting and potentially reckless endeavor.
what to know for later
🧑⚖️ Elon Musk challenges OpenAI's strategy. Elon Musk, via his startup xAI, has filed an injunction to prevent OpenAI's shift from its original non-profit model to a profit-driven structure. The legal action also targets OpenAI's alleged restrictions on investors funding competing firms, a move Musk claims disadvantages rivals like xAI. Read more
🎮 Genie 2: Big claims, limited scope. Google's Genie 2 generates interactive 3D environments from text or images, showcasing advancements in video generation and memory retention for AI models. However, its limitations in long-term consistency, real-time speed, and application viability highlight ongoing challenges in creating usable AI-driven virtual worlds. Read more
🎥 Tencent's open-source AI video model. Tencent's new HunyuanVideo, a 13-billion-parameter AI model, leads the open-source domain in text-to-video generation, offering high visual fidelity and scene dynamics. Key features include a video-to-audio synthesis module for realistic sound, scalable efficiency reducing computational costs by 80%, and advanced avatar animation tools. Read more
🎨 Aurora AI generator's brief debut. Aurora, an advanced image generator integrated into Grok on X, demonstrated high quality generations in depicting people and animals before being pulled hours after launch. It may have been removed for further development or to enhance safeguards against misuse, with speculation it could return as a top-tier AI tool. Read more