CyberFreezeDev
Loading
A neural network trained on huge amounts of text to predict what comes next.
An LLM predicts text one token at a time from patterns learned in billions of sentences. More parameters usually mean more ability and higher cost.
A large language model (LLM) is a program that has learned the statistical patterns of language from billions of sentences. Give it some text and it predicts what comes next, one token at a time. GPT, Claude, Gemini, Llama, Qwen and Mistral are all LLMs.
"Large" refers to the number of learned parameters, from a few billion in models that run on a laptop to hundreds of billions in hosted models. More parameters usually means better reasoning and more knowledge, and also more memory and cost to run.
Reading a huge body of text, the model learns grammar, facts and reasoning patterns.
The model reads a very large text corpus and learns to predict the next token. This is where it picks up grammar, facts and patterns of reasoning.
Training on good answers and human ratings teaches it to follow instructions.
It is then trained on examples of questions and good answers, and on human preference ratings, so it follows instructions instead of just continuing text.
You send instructions, context and a question, all inside the context window.
At use time you send a prompt: instructions, context and the question. Everything has to fit in the model's context window.
The model writes one token, appends it and repeats until done.
The model produces one token, appends it, and repeats until it finishes. This step is called inference.
Emails, chats, contracts and tickets summarised, tagged and routed without someone reading each.
Emails, WhatsApp messages, contracts and support tickets can be summarised, tagged and routed without someone reading each one.
Replies, quotes, product descriptions and reports, drafted for a person to check.
Replies, quotes, product descriptions and reports, drafted for a person to check and send.
With function calling, it can look up an order or create a CRM record.
With function calling, an LLM can look up an order, create a CRM record or start a workflow.
A CA firm's LLM labels each client email, pulls out the client and period, and drafts a reply that staff approve or edit.
A CA firm receives client queries by email all day. An LLM reads each message, labels it (GST filing, TDS, invoice copy, general), pulls the client name and period, and drafts a reply from the firm's own notes using retrieval-augmented generation. A staff member approves or edits the draft. The model does the reading and typing; the accountant keeps the judgement.
Built-in knowledge is frozen and can be wrong. Pass in facts from your own data.
An LLM's built-in knowledge is frozen at its training date and can be wrong. Facts that matter should come from your own data, passed in with the prompt.
Models can write fluent, confident text that is false. Design around it.
Models can produce fluent, confident text that is false. See hallucination for how to design around it.
A smaller open-weight model on your hardware often handles classification and extraction.
A smaller open-weight model running on your own hardware is often enough for classification and extraction, and keeps data in house.
ChatGPT is an application built on top of OpenAI's GPT models, which are LLMs. The app adds a chat interface, memory features, file handling and safety layers around the model.
Yes. Open-weight models such as Llama, Qwen and Mistral can run on a GPU workstation or server you control, so prompts and documents never leave your network. See our private AI deployment service.
Not during normal use. The model's weights do not change when you send a prompt. Whether a hosted provider keeps your data for training is set by its terms and your account settings, so check them before sending client data.
Last updated