What is a LLM
Definition
A Large Language Model (LLM) is an AI system trained on massive amounts of text to understand and generate human language.
All LLMs are AI models
But not all AI models are LLMs
How it works
LLMs learn by predicting the next token (word fragment) in a sequence. After training on billions of examples, the model picks up enough patterns to generalize to new inputs.
Key concepts
| Concept | What it means | Why it matters |
|---|---|---|
| Token | Small piece of text | This is how AI "counts" and it affets cost and limits |
| Context window | How much text the AI can “hold in memory" at one time | Limits how much of the conversation it can see at once |
| Harness | The system that wraps a model and makes it usable | Connects the model to tools, data, and APIs |
LLM models
| Model | Provider | Context window max | Best for |
|---|---|---|---|
| GPT-5 | OpenAI | 1M+ tokens | General reasoning, coding, multimodal |
| Claude Sonnet 4.6 | Anthropic | 1M+ tokens | Deep reasoning, long documents, analysis |
| Claude Opus 4.6 | Anthropic | 1M+ tokens | European data residency, multilingual |
| Gemini 2.0 / 2.5 | Google DeepMind | 1M+ tokens | General reasoning, coding, multimodal |
| Llama 4 | Meta | 200K tokens | Open-source apps, customization |
| Mistral Large 2 | Mistral AI | 200K tokens | Enterprise reasoning, coding |
Note: These are only example models and plans. AI models, pricing, and features change rapidly, so always check the official documentation for the most accurate and up-to-date information.
Limitations
| Limitation | Impact | Example |
|---|---|---|
| Hallucination | Generates plausible but incorrect facts | Students asked AI for scientific papers. It generated real-looking research titles, authors, and journals that were completely fabricated |
| Knowledge cutoff | Unaware of events after training date | An AI was trained in 2024. Someone asks it: “Who won the 2025 Olympics 100m final?”. It might give an outdated answer or guess incorrectly |
| No persistent memory | Forgets previous sessions | If someone chats with one model in the same session, the model will, at some point, compress the information and lose some context. |
| Context limit | Cannot process unlimited input | Someone pastes a 500-page document into an AI. Let's say it can only read, say, 100 pages worth of tokens at once so it will not be able to process everything. |
| No native actions | Cannot browse, run code, or call APIs alone | Someone asks an AI: “Book me a flight to Paris tomorrow”. It can describe how to book it, but it cannot actually book anything. |
Note: Some of these current limitations can be partially overcome by combining models with external tools, retrieval systems, memory stores, and agent frameworks, but the underlying model itself still relies on patterns learned during training and does not inherently gain guaranteed truth, live awareness, or autonomous capabilities.