Natural Language Processing (NLP) is the umbrella field for getting computers to work with human language — parsing it, classifying it, translating it, generating it. Large language models (LLMs) like GPT-class models are the most visible recent chapter of that field, but they didn't replace the fundamentals underneath NLP; they sit on top of them.
Pre-LLM NLP generally needed a purpose-built model per task. LLMs, trained on enormous amounts of text, can perform many of these tasks reasonably well with just a well-written prompt and no task-specific training — which is genuinely new. What they didn't add is reliability by default: LLMs can state incorrect things with complete confidence ("hallucination"), and they inherit biases and gaps from their training data. Generative AI systems that matter in practice are usually LLMs paired with retrieval, validation, or human review — not an LLM used blind.
Ask what the task actually needs. Classifying support tickets by urgency doesn't need a large generative model — a well-trained classical or lightweight model will be faster, cheaper, and more predictable. Drafting a first-pass response, summarizing a long document, or answering open-ended questions is where generative models earn their cost. Knowing the difference is a real, employable skill in itself.
The NLP, LLMs & Generative AI track in the Academy covers both layers deliberately — the classical NLP tasks that still power a huge share of production systems, and the generative techniques (prompting, retrieval, generation pipelines) that are newer but increasingly expected.
0 comments