NVIDIA GTC 2024
During Nvidia's GPU technology conference, CEO Jensen Huang presented several major announcements, highlighting Nvidia's advancements in artificial intelligence and computing:
1. Introduction of Blackwell, dubbed the world's most powerful AI chip, promising to outpace its predecessor (Hopper) in efficiency by 30 times and reduce energy and cost consumption by 25 times.
2. Unveiling of the GB200 Grace Blackwell Superchip and its availability on NVIDIA DGX™ Cloud. Key tech giants like Amazon, Microsoft, and Google are set to adopt the Blackwell system.
3. Omniverse's integration with Apple's VisionPro for advanced 3D development and streaming.
4. Launch of "Project GR00T," aimed at shaping general humanoid robotics, accompanied by the "Jetson Thor" computer for these robots.
5. Presentation of a 3D blueprint for AI-focused next-gen data centers, emphasizing their role as “AI factories.”
6. Expansion of collaboration with Oracle to deliver sovereign AI solutions enabling nations to bolster their digital sovereignty.
7. Rollout of a 6G research platform to foster the development of AI-influenced RAN technology essential for a hyper-connected world.
Grok Open-Sourced
On March 17, 2024, xAI open-sourced Grok-1, a colossal 314 billion parameter Mixture-of-Experts language model. Grok-1 is a base model, not fine-tuned for specific tasks, and was developed from the ground up using a custom-built training stack leveraging JAX and Rust, completed in October 2023.
The model's unique attribute is that only 25% of its expansive weight set is leveraged per token, optimizing its efficiency and capabilities. xAI has made Grok-1 openly available under the Apache 2.0 license, with setup instructions hosted on their GitHub repository.
Elon Musk delivered on his promise to open-source Grok, but only after suing OpenAI for being closed-source and being called a hypocrit.
Mustafa Suleyman, DeepMind and Inflection Co-founder, joins Microsoft to lead Copilot - Satya Nadella, CEO of Microsoft, announced the formation of a new division called Microsoft AI, aimed at driving innovation in consumer AI products and research, with a focus on their AI platform Copilot. Mustafa Suleyman, co-founder of DeepMind and Inflection, is appointed CEO of Microsoft AI and EVP, joining the senior leadership team and reporting directly to Nadella. Karén Simonyan, an influential AI researcher and Inflection's Co-founder and Chief Scientist, joins as Chief Scientist. The new team also includes experienced engineers and researchers from Inflection. This move emphasizes Microsoft's collaboration with OpenAI and its commitment to an AI-first strategy, with notable figures like Kevin Scott, CTO, continuing to oversee AI strategy. The organizational changes concentrate on enhancing Microsoft's consumer AI innovation, with various existing teams now reporting to Suleyman. Nadella underscores the potential of this initiative to produce groundbreaking technology that aligns with their mission to make AI benefits accessible and safe worldwide.
Apple and Google Are Discussing a Deal to Bring Generative A.I. to iPhones - Apple is in early talks with Google to potentially incorporate Google's generative AI model, Gemini, into future iPhones. Apple is also in discussions with other AI firms, reflecting its intent to integrate new AI features promised by CEO Tim Cook for this year. Competitors like Samsung and Google have already implemented Gemini into their devices for advanced functions such as video editing and audio summarization. The collaboration between Apple and Google, if finalized, would bolster a partnership dating back to the inception of the iPhone. The urgency for Apple to adopt AI technology comes as it lags behind in developing its large language model and loses its position as the world’s most valuable company to Microsoft.
Saudi Arabia Plans $40 Billion Push Into Artificial Intelligence - The Saudi government plans to launch a $40 billion artificial intelligence fund, signaling a major push into a transformative technology sector. In discussions with top Silicon Valley VC firm Andreessen Horowitz and others, this initiative would catapult Saudi Arabia to the status of the world’s largest AI investor. This move aligns with the nation's strategy to diversify its economy beyond oil and boost its global business influence via its sovereign wealth fund, which holds over $900 billion in assets. With the creation of the fund leveraging Wall Street banks' expertise, Saudi Arabia joins a competitive and cash-rich AI investment landscape, with firms like Anthropic recently securing substantial funding. This development also highlights the substantial costs associated with advancing AI technology, such as the funding of specialized chip manufacturing.
YouTube Requires Labeling AI - YouTube is implementing new labeling features for videos, aiming to enhance transparency regarding content that is altered or synthetically produced. These labels will display in the video descriptions, with more prominent placement for videos tackling sensitive subjects such as health and politics. The rollout will start with the mobile app, followed by desktop and TV platforms. Future enforcement actions are anticipated for creators who fail to declare the use of altered content. YouTube may also proactively label certain content to prevent misinformation. Additionally, YouTube is part of the C2PA coalition, further emphasizing its commitment to content authenticity.
Ex-Activision CEO Bobby Kotick pitched buying TikTok to potential partners, including Sam Altman - Former Activision CEO Bobby Kotick is reportedly considering purchasing TikTok amid congressional scrutiny over the app's Chinese ownership. Kotick has approached ByteDance's cofounder and OpenAI's Sam Altman as potential partners for this acquisition. The interest comes as a bill progresses through the US Congress that could force TikTok to either be sold to a US owner or face a shutdown. The House Energy and Commerce Committee has passed the bill, with a House vote pending; President Biden has signaled his support. TikTok, however, views the bill as a de facto ban, citing the practical challenges and effect on its global operations. Kotick, who encountered controversy during his tenure at Activision over alleged knowledge of workplace misconduct, and the other parties involved did not immediately comment on the matter. Lawmakers support the bill due to concerns over potential data sharing with the Chinese government.
Awesome Research Papers
This study delves into the optimization of Multimodal Large Language Models (MLLMs), emphasizing the critical nature of architecture components and pre-training data curation. The researchers conduct detailed examinations of elements such as the image encoder, the vision-language connector, and pre-training data varieties. Key findings indicate the significant impact of a balanced mix of image-caption pairs, text, and image-text data for achieving top few-shot learning performance. They discovered that while the image encoder's design and the quality of images and tokens are crucial, the vision-language connection is less influential. Leveraging these insights, the team developed MM1—a suite of up to 30 billion parameter multimodal models, including dense and mixture-of-experts variants—achieving state-of-the-art pre-training effectiveness and competitive benchmark results after supervised fine-tuning. MM1 models also showcase advanced properties like improved in-context learning and the capacity for multi-image reasoning, further enhancing their few-shot learning capabilities.
AutoDev introduces a novel AI-driven software development framework aimed at expanding the capabilities of traditional AI-powered programming assistants. Unlike its predecessors that offer limited functionalities, AutoDev provides a comprehensive suite of autonomous software engineering tasks ranging from code generation, testing, and execution to git operations, all within a secure Docker container environment. With a focus on user privacy and customizable operational boundaries, AutoDev has been evaluated on the HumanEval dataset, showcasing impressive results, with over 91% Pass rates in both code and test generation, thereby highlighting its potential to revolutionize autonomous software development.
