← AI Terminology

LLM - Large Language Model

A large language model is a neural network trained on massive text corpora to predict the next token — at sufficient scale, these models develop broad language understanding, reasoning, code generation, and instruction-following capabilities that emerge from scale alone.

Examples: GPT-4, Claude 3.5, Gemini 1.5, Llama 3, Mistral Large.
Why It Matters in AI
LLMs are the defining AI technology of the 2020s: they unified NLP tasks that previously required separate specialist models (translation, summarisation, QA, code) into a single general-purpose model. Scaling compute, data, and parameters predictably improves capability (scaling laws), enabling continuous improvement through larger training runs. Every major AI product — from ChatGPT to GitHub Copilot — is built on an LLM at its core.
Key Points
Aspect Description
Scale Hundreds of billions to trillions of parameters; trained on trillions of tokens
Fine-tuning Instruction tuning (SFT) + RLHF/DPO alignment converts raw LLM into a useful assistant
Architecture Transformer-based, decoder-only (GPT style) for generation; encoder-decoder for seq-to-seq
Pre-training Self-supervised next-token prediction on web-scale text — no human labels required
Frontier models GPT-4o (OpenAI), Claude 3.7 (Anthropic), Gemini 2.0 (Google), Llama 3.3 (Meta, open-weights)
Emergent abilities Capabilities like few-shot reasoning, chain-of-thought, and code generation emerge at scale
Simple Analogy
An extraordinarily well-read person who has absorbed the internet's worth of text and can continue any conversation, answer any question, or write in any style — not because they "understand" in a human sense, but because pattern completion at massive scale produces behaviour indistinguishable from understanding.
Common Usage Examples
  • OpenAI API: client.chat.completions.create(model="gpt-4o", messages=[{"role": "user", "content": "..."}])
  • Anthropic API: client.messages.create(model="claude-opus-4-7", max_tokens=1024, messages=[...])
  • Local inference: ollama run llama3 — run Llama 3 on consumer hardware
  • Fine-tuning: trl.SFTTrainer(model, train_dataset=dataset) — instruction-tune an open-weights LLM
  • Evaluation: MMLU, HumanEval, LMSYS Arena — standard benchmarks for comparing LLM capability
Summary
In short: Large language models are the transformer-based neural networks trained on internet-scale text that power modern AI — general-purpose systems capable of language, reasoning, and code at a level that has transformed the technology industry.