Inside the mind of LLMs

This is an upcoming series of articles about the workings of large language models (LLMs). Expect it to be fairly technical, but I will try to keep it accessible to people with a basic understanding of computer science, technology and mathematics.

Planned topics:

• Tokenizers, Tokens, Vocabulary
• Building Blocks: Query, Key, Value, Attention, Non-linearities, Layer Norm, Batch Norm
• The Evolution of Attention: From the "Attention is all you need" paper to the latest Qwen 4 architecture
• Logits and Next-Token Prediction
• Memory: KV Cache, MMap, FreeTokens
• Constrained Decoding
• Inference: Prefill, Decode, Batching, MTP, and hardware mechanics
• Quantization formats: FP16, BF16, FP8, INT8, INT4, NVFP4, and more
• Benchmarks: what they measure
• Teaching a model: SFT, instruction-tuning, domain-adaptation
• Fine tuning efficiently: full-finetuning, PEFT, LoRA, QLoRA
• Reinforcement Learning
• Distillation
• How reasoning works
• Vision transformers
• Diffusion models
• Emergent characteristics
• Key papers to read