According to the source, writing a traditional function means encoding deterministic rules: for each possible input the programmer specifies exactly what the code should do. This approach works well for tasks such as sorting a list, calculating a price or processing a payment, where every relevant case can be anticipated and an instruction written for it. The article notes that this method breaks down when the problem requires interpreting human language, for example deciding whether a customer support message expresses frustration. Trying to cover every possible phrasing — such as ‘angry’, ‘ridiculous’ or an indirect comment like ‘I have been waiting three weeks’ — quickly becomes infeasible because language cannot be reduced to a finite set of conditions. The source explains that large language models (LLMs) operate differently. Rather than following explicit rules, an LLM is a prediction machine that has been shown millions or billions of text examples. During training the model discovers patterns in those examples, and when it receives a new input it predicts the most likely continuation based on what it has seen. The article emphasizes that an LLM does not know facts with certainty; it generates the most probable next sequence of words, which is why it can sound confident while actually guessing. A further characteristic highlighted is that LLMs are stateless. Each request is processed in isolation, with no retention of prior messages. The illusion of memory in tools like ChatGPT or Claude arises because the application sends the entire conversation history back to the model on every call, effectively providing context. As an AI engineer, one must curate this context, deciding what the model can and cannot see. The source defines context as everything the model can see at the moment it generates a response: the system prompt, the conversation history, any relevant documents, and the user’s current message. If information is outside this set, the model does not know it exists. The amount of context that can be passed is limited by the context window, measured in tokens. The context window is the maximum amount of text an LLM can process in a single call; all components of the context must fit within it, and anything beyond is invisible to the model. Consequently, optimizing the use of the context window shapes many design decisions in AI engineering.
