In plain words · technology · unofficial
Claude in plain words: what this AI is, how Anthropic trains it and what happens between your question and its answer. Nine steps, an unofficial walkthrough based on Anthropic’s public materials. The diagram plays by itself; drag the scale below it.
In short
Claude is a language model made by Anthropic. It was trained on a large body of text to predict the next word and then further trained to be helpful and honest. When it answers, it reads the whole conversation as context, can think before replying and can search the web when needed.
Marcella, web and bot developer · updated 17 November 2026
Step 01 / 09
First Anthropic trains the model on a huge amount of text and code: public web pages, data from third parties and labelers, chats of users who allowed training on them, and data the company creates itself. At this stage there is one task — guess the next piece of text.
Technically: this gives a base model. It already knows the language and many facts but cannot yet hold a conversation like an assistant.
Claude is an AI assistant and a family of language models from the American company Anthropic. The company was founded in 2021 by former OpenAI employees; the first version of Claude was shown in March 2023. Claude writes and edits text, answers questions, works through documents and images, writes program code. It runs at claude.ai, in desktop and mobile apps and through the API.
Inside Claude is a large language model, as in other chatbots: it writes the answer one token at a time, and its knowledge and skills live in its weights. First the model is trained to continue text on a huge dataset, then fine-tuned to be a helpful and safe assistant. Anthropic calls its approach “Constitutional AI”: the model checks its answers against a set of principles that the company publishes openly.
There are several models of different sizes — for example Haiku, fast and inexpensive; Sonnet, an all-rounder; and Opus, for hard tasks. New versions come out several times a year. The models can reason before answering, read up to a million tokens at once and use tools: search the web, run code, work with files and connected services.
Keep the limits in mind. Like any neural network, Claude can be confidently wrong, and it knows about events after its training date only if it is given search. By itself it does not remember conversations: memory in the app is notes added to the context. This is an unofficial walkthrough: I am not affiliated with Anthropic and write from its public materials.
I will pick a model for your task, budget and the country your service runs in, and build it into a Telegram bot or website — like my “Что там?” bot, where a neural network builds a topic profile from a couple of lines.