Claude and Anthropic Basics
Claude is a family of AI assistants / Large Language Models (LLMs) built by Anthropic. Anthropic was started in 2021 by Dario Amodei, Daniela Amodei and a few former OpenAI researchers. The company's focus is AI safety: building powerful AI that is helpful, honest and harmless (HHH).
1.1 What is an LLM?
An LLM is a large neural network (Transformer architecture) trained on a huge amount of text. Its basic job is to predict the next token after the given text. Scaling this one simple task is what produces abilities like reasoning, coding, summarization and translation.
- 1.1. Prompt: the user's text / image / PDF
- 2.2. Tokenizer: split the text into tokens
- 3.3. Model: Transformer predicts the next token
- 4.4. Sampling: pick a token, repeat
- 5.5. Response: tokens turned back into text
The model generates one token at a time until a stop condition is reached
1.2 Key terms
| Term | Meaning | Example / Note |
|---|---|---|
| Token | A small piece of text (part of a word) | In English ~1 token ≈ 3-4 characters; Hindi/Hinglish uses more tokens |
| Context window | How many tokens the model can see in one request (input + output) | Current Claude models: 1M tokens (Haiku 4.5: 200K) |
| Max output | Maximum tokens in a single response | Newer models: up to 128K (streaming is required for large outputs) |
| Prompt | The input given to the model | System prompt + messages |
| System prompt | An instruction that sets role, rules and context | "You are a senior .NET reviewer..." |
| Inference | Getting output from a trained model | Every API call is one inference |
| Hallucination | The model states something wrong with confidence | Ask for sources/quotes, give permission to say "I don't know" |
| Knowledge cutoff | The date up to which the training data goes | For later information, provide web search / docs |
| Multimodal | Understanding more than text, such as images/PDFs | Claude can read images and PDFs (input); output is text |
| Agent | An LLM that works in a loop using tools | Claude Code is a coding agent |
1.3 How Claude is trained
- 1.1. Pre-training: next-token prediction on huge data from the internet, books and code
- 2.2. Supervised fine-tuning: learning to behave like an assistant from good example conversations
- 3.3. RLHF: Reinforcement Learning from Human Feedback — humans rank responses
- 4.4. Constitutional AI / RLAIF: the model critiques and improves its own answers based on principles (a constitution)
- 5.5. Safety testing: red-teaming, evals, Responsible Scaling Policy checks
Constitutional AI is Anthropic's signature technique
Constitutional AI (CAI): Anthropic introduced this method in 2022. Humans do not label every harmful example; instead the model is given written principles (a "constitution") and it critiques and revises its own answers. Reinforcement learning then runs on AI feedback (RLAIF). Anthropic has also published Claude's constitution publicly, which sets out Claude's values, character and priorities.
1.4 What makes Claude stand out
- Long context: can read a whole codebase, long contracts or several PDFs at once (1M tokens).
- Coding and agentic work: Claude is well known for coding benchmarks and long-running agent tasks (Claude Code is built on this).
- Tool use: calling functions/APIs, web search, code execution, computer use.
- Extended / adaptive thinking: reasoning internally before answering, for better accuracy on hard problems.
- Safety: refusing when a request is harmful, staying honest, stating uncertainty.
- Character: curious, warm, direct — trained to avoid sycophancy (just agreeing with everything).