Separate the four layers before you use the words

A chat can write, answer and rephrase; that does not make AI, a model, ChatGPT and an agent synonyms.

“It sounds true” is generation; “it can be proven true” requires sources, retrieval and verification.

01 / VOCABULARYSeparate the four layers first

AI is the largest umbrella: machines performing work that once required human judgement. Machine learning is one way to do it, learning patterns from data. A large language model is a machine-learning model for sequences such as language and code. An agent places a model inside a work system that can observe, act and verify. AI includes recommendation, vision, speech and autonomous driving. ML learns rules from data. LLMs predict continuations. Agents add goals, state, tools, permissions, verification and stopping conditions.

02 / GENERATIONGeneration starts with what comes next

In a simplified view, a language model receives context, divides text into tokens, and predicts the next token from patterns learned in training. The new token joins the context and the process repeats, producing an answer step by step. This is not mechanical repetition: at scale, next-token prediction learns grammar, conceptual relations, common reasoning patterns and extensive knowledge. Yet its objective is a plausible continuation, not an automatic guarantee that every sentence is true.

03 / BOUNDARIESKnow strengths and the ways it fails

Models are useful for rephrasing material, finding structure in much text, proposing alternatives, applying explicit rules to a particular document and producing drafts. They can invent facts, miss recent changes, misread an ambiguous request or give a complete answer with insufficient evidence. Verify dates, prices, legal rules, medical advice, people’s status, citations and calculations. Never directly delegate irreversible actions such as payment, deletion, public publishing, external sending or permission changes.

04 / FIRST TRYRun a minimal context experiment

Choose a small question you know. Ask it once with no more context; then add the use case, existing material, output format and an instruction to mark uncertainty. Compare the results: does the second better meet the goal, and which facts still need your own verification? This is the smallest context experiment. You have not built an agent, but you have begun to observe model behavior as an engineer.

Next: what happens inside a chat.

Continue to chat systems

Separate the four layers before you use the words.