How LLMs Work✎

The Transformer is a neural network architecture built entirely on self-attention and feedforward layers, introduced in the 2017 "Attention Is All You Need" paper, and the foundation of virtually all modern LLMs.

LLMs generate text by iteratively predicting one token at a time, converting final-layer vectors into probability distributions over vocabulary via an unembedding matrix and softmax, with temperature and decoding strategies controlling the creativity–coherence tradeoff.

LLMs represent language as high-dimensional numeric vectors, converting raw text into tokens and then embeddings that encode semantic relationships geometrically — the foundation on which attention and all downstream computation operate.

Self-attention is the core computational primitive of the Transformer, allowing every position in a sequence to directly attend to every other position by computing query, key, and value vectors and weighting their contributions via scaled dot-product scores.

Training a modern LLM happens in two main stages — expensive self-supervised pretraining on massive text corpora followed by cheaper supervised finetuning and optional RLHF — with performance scaling predictably as a power law of parameters and data.

Beyond text generation, LLMs are increasingly framed as general-purpose reasoning kernels that coordinate tools, memory, and multimodal inputs — but they remain System 1 thinkers, constrained by context windows, unidirectional knowledge, and the absence of robust deliberate reasoning.

RLHF aligns LLMs with human preferences by training a reward model on ranked human comparisons and then optimizing the language model policy against it via proximal policy optimization, with DPO and related direct alignment algorithms offering simpler alternatives.

RAG extends LLMs by retrieving relevant documents from an external knowledge base at inference time and injecting them into the prompt, grounding responses in up-to-date or domain-specific information without retraining the model.

Mixture of Experts is an ensemble architecture where a learned gating function routes each input to a subset of specialized sub-networks, enabling massive parameter counts while keeping per-token compute tractable — a key strategy for scaling modern LLMs.

Fine-tuning adapts a pretrained model to a downstream task by continuing training on new data, with parameter-efficient methods like LoRA and adapters enabling this at a fraction of the cost by updating only a small subset of weights.

As LLMs are deployed as agentic systems that read external data and take actions, they become vulnerable to jailbreaking, prompt injection, and data poisoning — three distinct attack classes that exploit the model's inability to distinguish trusted instructions from adversarial inputs.