If you work with large language models, whether you're building applications, fine-tuning, or using them through an API, you're working with transformer models. And a lot of great resources have been created to explain how they work. Courses that build them from scratch, complex explanations of the attention mechanism, and theoretical deep dives. But you might find that those explanations don't always connect to the problems you actually face using transformers. This course connects all the critical pieces to building and using transformer networks in practice. It builds a coherent picture from how a trained transformer generates text, to what's happening during a forward pass of the model, to how the process gets optimized to run on GPUs. I'm delighted to introduce Sharon Zhou, who's your instructor for this course on transformers in practice. I've known the work of Sharon for many years, starting from her days at Stanford, and we've really stayed friends and kept in touch through her now being VP of Engine AI at AMD. So Sharon, say a bit more about why you built this course. Yeah, absolutely. The goal of this course is really to help you, the learner, gain deeper intuition in a way that impacts how you can practically use LLMs, not just the fun math part. I've noticed, you know, people often hit on some issue when they're working with language models, and if only they had that deeper intuition, and they could actually debug it instead of say, well, LLMs don't work. I think that once you understand the core of what these models do, generating text one token at a time from a probability distribution, everything else follows. Now, one interesting evolution of the transformer network is the original Google brain paper on the transformer network talked about the encoder-decoder architectures, and one of the most important evolutions has been that all the frontier labs now pretty much use the decoder-only architecture. Yeah, in this course, the focus is really on decoder-only transformers, because that's the architecture behind nearly all the frontier models that you're using today. So you'll go deep on practical concepts that matter most for your work. And throughout this course, there are some also really fun, informative, interactive visualizations where you can actually see these processes in action. Like you can watch tokens get sampled, watch attention scores form, see what's happening inside the GPU. I found that playing with the concepts helps with gaining deeper intuition on them. Yeah, I think the concrete examples are important, because when it comes to transformers, there's often more than meets the eye. Let's walk through as well what the course actually covers. Okay, let's do it. So in module one, you learn about everything an LLM does, whether it's writing code or why it's hallucinating. And that all comes from the same process, picking one token at a time from a probability distribution of all the possible tokens in its vocabulary. This basically helps explain many LLM behaviors that fall out of this process and shows how techniques like structured output or RAG, retrieval augmented generation, and reasoning all work by building on that same loop. And that raises an obvious question, where does that probability distribution even come from, right? That's in module two. So you'll look inside the model there, the attention mechanism, how positional information is encoded and included, how layers build on each other to see the math that produces it. It's a lot of matrix operations and seeing them helps you understand what the model is actually doing with your input. Finally, once you see the math involved in a single forward pass, you can begin to understand what goes on in a full LLM system beyond just the language model itself. So in module three, you'll look at how that math runs on GPUs, the optimizations like KB caching, flash attention, quantization. Again, don't worry if you don't know what those terms are. These are techniques that make these larger frontier models fast enough to actually serve users. One of the things I really like about this course is that if you're already working with LLMs, you probably have a sense of how they behave from your practical experience, but not necessarily always why they behave that way. By the end of this course, you have that coherent picture that Sharon described that will help you to make better decisions about deployments or help you to follow new research and really understand what's happening when things go right or when they don't go right. I hope you enjoy learning about how transformers work and let's go on to the next video and get started.