• AI Emergence
  • Posts
  • Everything @Nvidia GTC 2024 and craziest hiring ever!

Everything @Nvidia GTC 2024 and craziest hiring ever!

Is Siri About to Ask, 'Hey Google, What's Up?

Hey there, 

“In the past 8 years, AI computing has grown 1000x!" - Jensen Huang

It's true – the incredible power driving LLMs like ChatGPT, Gemini, and cutting-edge chatbots transforming our world often hides in the shadows. We rarely focus on the underlying hardware that makes it all possible.

But this week it all came to life during the Nvidia GTC 2024. With the unveiling of the new Blackwell architecture, we're set to witness an even greater level of generative AI with dramatically improved efficiency.

(Speaking of events, if you're passionate about GenAI, don't miss India's biggest and most awaited GenAI event of the year – The DataHack Summit. It's your chance to stay ahead of the curve!).

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection 

  • Question to ponder before we meet next!

Table of Contents

Recently, Nvidia organized a three-day conference focused on unveiling groundbreaking announcements in AI technology from Nvidia. The highlights included:

Nvidia Blackwell

Blackwell, the next-generation architecture for building high-performance AI chips ensuring better AI computing performance, and better energy efficiency. 

  • The latest B200 GPU boasts 20 petaflops of FP4 power with its 208 billion transistors.

  • Pairing these two GPUs with a Grace CPU in the GB200 configuration boosts LLM inference workloads by up to 30x as compared to the same number of H100.

  • This setup could reduce costs and energy consumption by as much as 25 times compared to the H100. (source)

Nvidia DGX B200

The DGX B200 is built on the cutting-edge Blackwell GPU architecture, optimized for generative AI and high-performance computing.

It houses up to eight GB200 Grace Blackwell Superchips, each combining a powerful CPU and GPU for incredible AI processing capabilities, promising significant performance improvements.

With the capacity for 3X faster model training and 15X quicker inference than its predecessors, the DGX B200 is equipped to handle a range of workloads, from large language models to chatbots. (source).

Other updates: 

  • Nvidia Microservices

    NVIDIA microservices provide easy-to-deploy software containers for generative AI and accelerated computing tasks, available through various cloud providers and vendors.

    Examples include NVIDIA NIM for optimized model inference and CUDA-based microservices for tasks like RAG, guardrails, and high-performance computing (source).

  • Nvidia GR00T

    Project Groot introduces a foundational AI model to streamline industrial robot development by learning from human observation, aiming to improve robot coordination and real-world interaction.

    This is supported by NVIDIA's new Jetson Thor SoC with the Blackwell architecture, delivering powerful AI capabilities crucial for running models like GR00T, alongside updates to the Isaac robotics platform. (source)

  • Omniverse

    Highlighting the fusion of AI with physical reality, Omniverse Cloud APIs deliver advanced simulation capabilities essential for the development of robotics and AI’s understanding of our world. (source)

Microsoft’s hiring of Inflection AI Leaders to run Microsoft AI

Satya has pulled it off one more time!

Microsoft has made a major move in the AI sector by hiring the co-founders of Inflection AI, Mustafa Suleyman, and Karén Simonyan, alongside key members of their team.  Suleyman, renowned for co-founding DeepMind (later acquired by Google), will now lead Microsoft's new consumer AI division. 

Consider the fact that this strategic hire comes less than a year after Microsoft led a significant funding round for Inflection AI. Additionally, Reid Hoffman, co-founder of Inflection also sits on Microsoft’s board. This is a second master stroke in less than 6 months from the Master! I am also noticing the pattern of investment and then hiring of founders and wondering what’s up with Mistral 😃 

Suleyman will oversee AI development for products like Bing, Edge, and Copilot. Meanwhile, Inflection AI announced a pivot in its strategy. The company will now specialize in building and testing custom generative AI models for clients.  They will leverage Microsoft Azure to expand their reach. (source)

  • Introducing Apple’s MM1

    Apple researchers recently published a paper titled "MM1: Methods, Analysis & Insights from Multimodal LLM Pre-training". This paper details their innovations in creating multimodal large language models (LLMs).

    Although not yet released, MM1 has potential applications such as improving Apple's Siri to answer questions based on images, showcasing its advanced capabilities. (source)

  • Apple may tap into Google's advanced AI for next-gen iPhone features

    Apple is in discussions to incorporate Google's AI platform, Gemini, into the iPhone, which could significantly expand the AI features in the upcoming iPhones.

    These "active negotiations" aim to license Gemini's AI models for generative AI features in an upcoming iPhone software update. This follows Apple's previous talks with Microsoft-backed OpenAI.

    However, the deal has not been finalized, and Apple might consider other AI providers like OpenAI or Anthropic if negotiations fall through. (source)

  • Apple secretly acquires DarwinAI

    Apple Inc. has quietly acquired the Canadian AI startup DarwinAI, marking a significant move to bolster its capabilities in generative AI for 2024.

    This acquisition has seen dozens of DarwinAI’s employees integrating into Apple's artificial intelligence division, according to insiders. The deal's specifics remain undisclosed as it has not been officially announced yet. (source)

XAI has made Grok AI open source under the Apache 2.0 license, introducing a 314 billion parameter Mixture-of-Experts model, Grok-1. This chatbot blends advanced machine-learning technologies for rich, engaging user interactions.

The GitHub link Grok AI’s open-source project offers details on using the Grok-1 model, including setup instructions and requirements. (source)

Google AI introduced Cappy, a compact pre-trained scorer model designed to boost the performance of large multi-task language models (LLMs) while being more efficient.

Unlike traditional LLMs that require significant computational resources, Cappy aims to make training and inference more cost-effective and accessible.

It operates as an independent classifier or as an auxiliary to LLMs, enhancing their performance without the need for extensive fine-tuning or access to their parameters.

Cappy's development addresses the challenges of adapting LLMs to complex tasks by offering a scalable solution that could streamline the use of AI in various applications. (source)

Amazon has introduced new generative AI features to assist sellers in creating high-quality product listings with less effort. By leveraging GenAI, sellers can transform existing product pages from their websites into rich, engaging listings on Amazon.

This innovation reduces the time and effort needed for sellers to generate compelling product descriptions, titles, and details, significantly improving the quality of AI-generated content. (source)

Adobe's AI image creation tool, Firefly, has made similar controversial mistakes as Google's Gemini, notably in producing inaccurate racial and ethnic depictions.

Despite Adobe and Google's different data training approaches, both encountered issues with their AI systems generating historically incorrect images. 

This highlights the industry-wide challenge of balancing the desire to avoid racist stereotypes with historical accuracy, a task complicated by the technology's inherent limitations and the diverse aims of AI model designers. (source)

Google DeepMind's SIMA project aims to develop AI agents that can understand and execute natural language instructions across diverse 3D virtual environments. 

This step towards artificial general intelligence (AGI) integrates language with action, allowing agents to adapt and perform tasks in complex settings. 

By focusing on language-driven generality and using behavioral cloning from human gameplay, SIMA demonstrates significant potential but also reveals challenges, particularly in intricate game environments. 

The project is a collaborative effort, emphasizing ethical development and the responsible use of AI technology. (source)

Tool: Tome

Problem Statement:  Quickly generate a basic presentation draft when you have a clear idea of the structure, but lack the specific content and a polished design aesthetic.

Use Case: Creating a startup funding deck for "XYZ"

Prompt - XYZ is an online DTC store that sells awesome designs created by GenAI. We sell it for INR 700 and we have a 50% margin. Our audience is people from Tier 1 cities aged from 18 to 35. We have an incredible team of fashion enthusiasts and our advantage is that we come up with new designs every week.

Overall Assessment:

Pros:

Ease of Use: An intuitive interface simplifies the initial draft creation process, especially when the overall presentation concept is established.

Design Variety: Offers a range of design options, allowing for basic customization of the draft's look and feel.

Cons:

Image Quality: Generated images may appear cartoonish or overly simplistic, potentially detracting from a professional business presentation.

Content & Design Refinement: Likely requires further editing and design adjustments to achieve a polished, investor-ready deck.

  • There has been an overwhelming number of AI tools released in the last few months. This blog post by Huyen Chip provides a comprehensive analysis of the current open-source AI ecosystem, focusing on the surge in repositories related to foundation models.

  • Check out how the Devin AI agent autonomously started a Reddit thread offering website-building services, cleverly navigated challenges, and even decided to charge for its work, underscoring the complex potential of AI agents.

  • Check out this interesting podcast featuring Sam Altman and Lex Fridman, where Sam Altman discusses the recent launches of OpenAI. This conversation covers topics ranging from lawsuits involving Musk to discussions about competitors like Gemini. I'm sure you're going to find it Interesting.

  • This new course Efficiently Serving Large Language Models by Deeplearning.ai builds a ground-up understanding of how to serve LLM applications.

  • In the recent episode of "Leading with Data," I had this conversation with Rajat Monga about his visionary role in shaping modern machine learning with transformative AI initiatives, including his leadership in the evolution of TensorFlow at Google Brain, OpenAI's approach to open source, his passion for research and product innovation at Inference.io, exploring the future of computation and Generative AI.

The discussion between Lex Fridman and Sam Altman was the highlight for me this week (outside of Nvidia GTC). Sam’s thought to think about computing as we think about Energy and not as hardware (at 1:10:00 timestamp) is profound in multiple aspects!

For example, in our analogy, you choose from 2-3 different classes of compute (5 Amp, 15 Amp, Commercial Switches) and just get charged on the number of FLOPs you used irrespective of the number of machines, server sizes, and other parameters. How would that change the design of the applications and systems?

The other profound statement is that he expects “Slow thinking” to take higher compute than “Fast thinking”. What do you think?

How do you rate this issue of AI Emergence?

Would love to hear your thoughts

Login or Subscribe to participate in polls.

Reply

or to participate.