CookLLM: Go deep. Build real things.

Hey everyone! I'm Harry 👋 I spent years as a computer vision engineer before switching to LLMs. The transition was painful — scattered blog posts, dense papers, and tutorials that either oversimplify or skip the hard parts entirely. So I built CookLLM, the course I wished existed. Here's what makes it different: 🔧 Everything from scratch — no HuggingFace wrappers, no black boxes. You write the tokenizer (BPE in Rust), the model architecture, GPU kernels in CUDA/Triton, Flash Attention, the full pretraining pipeline, and eventually SFT/RLHF. 📊 Interactive visualizations — concepts like RoPE, attention patterns, and memory optimization come with hands-on diagrams you can play with, not just static images. 🧠 Systems engineering included — most courses skip GPU programming and memory optimization. We don't. You'll write Triton kernels and understand why Flash Attention works at the hardware level. I'd love to hear your feedback — what topics would you want to see covered first? Happy to answer any questions about the curriculum or the from-scratch approach!