Large Language Models

Next-token prediction, scale, emergent behaviour, and a clear-eyed account of capabilities and limits.

Assumes: Encoder, Decoder and Encoder–Decoder Families

A modern LLM block dissected: pre-norm, RMSNorm, SwiGLU, GQA and the residual stream.

Assumes: What an LLM Actually Is · Transformer Shapes: End-to-End Walkthrough

Causal language modelling, the loss over a batch, teacher forcing, and packing sequences efficiently.

Assumes: Decoder-Only Architecture in Detail

Sourcing, deduplication, quality filtering, decontamination and data mixtures.

Assumes: Pretraining Objectives

Kaplan and Chinchilla power laws, compute-optimal token budgets, and worked allocation arithmetic.

Assumes: Pretraining Data Pipelines

FLOPs per token, the 6ND rule, GPU-hour estimation, and inference cost per million tokens.

Assumes: Scaling Laws

Rotary embeddings derived, relative position bias, and how context extrapolation is achieved.

Assumes: Positional Encoding

The quadratic cost of attention, long-context techniques, and what actually degrades at long range.

Assumes: RoPE, ALiBi and Positional Schemes

Why generation is memory-bound, cache size arithmetic, prefill versus decode, and paged attention.

Assumes: Context Windows

Sharing key/value heads, IO-aware attention kernels, and sparse and linear attention variants.

Assumes: The KV Cache

Sparse routing, top-k gating, load balancing losses, and the active vs total parameter distinction.

Assumes: Efficient Attention: MQA, GQA and FlashAttention

Greedy, beam, temperature, top-k, top-p and min-p, with the same logits decoded every way by hand.

Assumes: Pretraining Objectives

Draft-and-verify acceleration, acceptance rates, and medusa/lookahead variants.

Assumes: Decoding Strategies

Instruction clarity, role and delimiter use, few-shot exemplar selection, and output format control.

Assumes: Decoding Strategies

Eliciting intermediate steps, zero-shot CoT, self-consistency voting, and where CoT fails.

Assumes: Prompt Engineering Fundamentals

ReAct, tree-of-thought, least-to-most, self-refine and program-aided prompting.

Assumes: Chain-of-Thought Reasoning

What happens mechanistically when a model learns from the prompt, and induction heads.

Assumes: Chain-of-Thought Reasoning

JSON schema enforcement, constrained decoding, grammars, and tool-call protocols.

Assumes: Prompt Engineering Fundamentals

Where SFT sits between pretraining and preference optimisation, and what each stage contributes.

Assumes: Pretraining Objectives

Learning from pairwise preferences, the Bradley–Terry model, and reward hacking.

Assumes: Supervised Fine-Tuning: The Alignment Pipeline

The full RLHF loop, the KL penalty against the reference policy, and its practical instabilities.

Assumes: Reward Modelling · TRPO and PPO

Deriving DPO from the RLHF objective, plus IPO, KTO, ORPO and SimPO compared.

Assumes: RLHF with PPO

Replacing human labels with model-generated critiques against an explicit set of principles.

Assumes: DPO and Direct Preference Optimisation

Long chain-of-thought training, RL on verifiable rewards, and trading inference compute for accuracy.

Assumes: DPO and Direct Preference Optimisation

Why next-token prediction fabricates, calibration and uncertainty, abstention, and grounding strategies.

Assumes: Reasoning Models and Test-Time Compute

MMLU, GPQA, HumanEval and friends; contamination, saturation and why leaderboards mislead.

Assumes: Hallucination

Model-graded evaluation, position and verbosity bias, rubric design, and agreement with humans.

Assumes: LLM Evaluation and Benchmarks

Jailbreaks, direct and indirect prompt injection, data exfiltration risks, and layered defences.

Assumes: Structured Output and Function Calling

Continuous batching, throughput vs latency, tensor parallelism, vLLM and capacity planning.

Assumes: The KV Cache

The perceive–plan–act loop, tool use, memory, reflection, and honest failure modes.

Assumes: Advanced Prompting Patterns · Informed Search and A*

Role specialisation, debate, orchestration topologies, and when multi-agent is worse than one good prompt.

Assumes: LLM Agents

Vision encoders, projection into token space, interleaved training, and audio and video extensions.

Assumes: LLM Agents · Self-Supervised and Contrastive Learning