3/ Next steps: 1. Write the training loop 2. Get it actually learning 3. Then clean it up and make it more production-ready This “make it work first” approach is way more fun and less overwhelming.

3/ What I’m planning to cover: • Modern architecture blocks (MoE, RMSNorm, SwiGLU, MLA…) • Efficient pretraining • Fine-tuning techniques • Moving from notebooks to more production-style training…