AI Weekly Digest #15: AMD joins the AI GPU race!

Did Meta just release the future AI models?

June 18, 2023

AI Weekly Digest #15: AMD joins the AI GPU race!

Hello, tech enthusiasts! This is Wassim Jouini and Welcome to my AI newsletter, where I bring you the latest advancements in Artificial Intelligence without the unnecessary hype. Let's dive into this week's news and explore the practical applications of AI across various sectors.

  • The Main Headlines: AMD joins the AI race, OpenAI announcements, Google releases a new model for fashion & shopping, and finally we have an “AI generated voice” detector!

  • Beyond the Hype: Did Meta just release the future of AI models?

  • Bonus: How to build your own AI generated QR code?

Beyond The Hype

One of the primary criticisms of modern AI models, such as ChatGPT is their lack of internal knowledge representation.

As Yann LeCun frequently emphasizes, these large models are essentially auto-regressive models whose main function is to predict the next token or word. To build a truly intelligent model, a different architecture is needed - one that enables internal representation from which the model can make inferences.

This is a fundamental component of I-JEPA, the new model that has been open-sourced by Meta!

The I-JEPA model can extrapolate based on a portion of an image and its internal knowledge representation. For instance, it knows that it’s a dog before creating the rest of the image.

What should you know about this new architecture?

  • Simulates Human Learning: I-JEPA is special because it tries to learn in a way that's similar to how humans do. When we learn, we create pictures in our minds of how things work. I-JEPA does something like that but for AI- it builds an internal understanding of images.

  • More Efficient: What’s really cool is that I-JEPA is compute efficient as it requires less examples to reach the same results!.

  • Open to Everyone: Both the model and the dataset are open source! You can also think about how you could leverage this new architecture to build the next generation of AI models!

Main Headlines

As the GPU shortage continues to impede the development and deployment of AI applications, AMD is making a significant entrance by announcing

  • The MI300X, their latest GPU, which is presented as the most powerful to date, even compared to Nvidia’s H100, and optimized for AI applications. Excitingly, the MI300X is capable of running large-scale open-source models, such as Falcon 40B, on a single chip. The MI300X is expected to hit the market by Q3 2023!

  • AMD x HuggingFace: Additionally, AMD is forging a partnership with Hugging Face, the largest platform for open-source AI models and datasets. This collaboration entails integrating AMD's GPUs, including the upcoming MI300X, with the Hugging Face platform, enabling users to leverage AMD's hardware for running an extensive range of models available on Hugging Face.

Main 3 announcements to remember:

  • Cheaper gpt3.5-turbo (4k models): 25% price drop effective immediately!

  • A new gpt3.5-turbo 16k context window model! This model comes with:

    • A slightly more expensive pricing (~2x more expensive than the 4k version)

    • BUT it enable you to handle up to 16k tokens (~12-15 pages!) enabling scenarios that were impossible to handle previously at reasonable price!

  • Function calling: OpenAI's recent versions introduced a new feature that allows users to define functions while interacting with GPT models. If a specific function is required while calling the model, the model returns a "function_call" parameter, indicating the need to call a designated external function to obtain a desired output.
    Why is it important? you’ll be able to design full systems relying on both the GPT model and external APIs. The GPT model playing the role of coordinator indicating explicitly which function to call, and with what parameters, whenever it’s required!

Disclamer: virtual try-on models proved to be challenging, as it requires to perfectly match the clothes, you want to try, on the target body and shape.

What’s new? get the convenience of trying on clothes virtually and finding the perfect fit and style without leaving your home. With this new release, you can see how clothes look on a diverse range of real models, representing different skin tones, body types, and hair types. This generative AI tool accurately reflects how the clothes drape, fold, cling, stretch, and form wrinkles and shadows on the models.

ElevenLabs has introduced an AI Speech Classifier that allows users to upload audio samples to detect if they contain audio generated by ElevenLabs' AI, aiming to promote transparency and responsible use of AI-generated content, a move that comes in response to widespread criticism that the company has made voice cloning too easy.

Bonus!

The AI-generated QR code that has been making waves recently!

QR code in the form of a Dog in front of a wooden wall. Look at how the QR code elements & colors adapts to the Dog’s shape!

If you're intrigued by QR codes integrated into AI-generated images, learn how to make your own by following this tutorial: https://qrdiffusion.com/tutorial/generate-qr-codes-with-ai.

Note: These QR codes require a dedicated QR code scanner! Smartphone cameras, which typically try to detect QR codes, often struggle with these.

That’s it for today! If you made it this far, I’d appreciate a quick feedback 😋! I know there’s room for improvement! So don’t hesitate to share with me the things you liked and those that you didn’t.

Have a great Sunday and may AI always be on your side!