Andrew introduces a new course on Mistral's 8x7B open-source model, which has 46.7 billion parameters but activates only 12.9 billion per token. Sophie Lian from Mistral AI covers model selection, prompting, function calling, and RAG in the course taught by Silver Yang.
I'm delighted to introduce Getting Started with Mistral, built in partnership with Mistral AI and taught by Silver Yang. One of Mistral's exciting open source models that you see in this course is the Mistral 8x7B model, which improves on the standard transformer neural network architecture by using a mixture of XSPLIT architecture that includes eight XSplit networks, that is eight distinct feedforward neural networks, so that at inference time for each token a separate gating network chooses two of these Xplus Activate, and it finally uses a weighted average of these two networks to choose the next token. This mixture of experts architecture enables a model that has both the improved performance and benefits of a larger model, both the inference speed of a much smaller model. Concretely, even though mixture has 46.7 billion parameters, because it activates only two of the eight expert networks to predict each token, it only uses 12.9 billion of its parameters at inference time to predict that next token. I'm delighted to introduce the instructor, Sophie Lian, who is head of Developer relations at Mistral AI, who regularly teaches developers about generative AI on YouTube and at conferences.
Thank you Andrew. In this course you will access Mistral's collection of open source and commercial models, get some advice and intuition on when to use each one depending on your tasks, and also get familiar with some of the useful features including prompting function, calling, JSON mode and rag.
By the way, le Mistral is a French word meaning strong northerly win, so I think with Sophia's help you breeze through this course on this draw.
