NLP Myths

Language AI attracts more hype than almost any other technology. Headlines swing between it's conscious, and it's useless. Neither of those is true. Let's clear up the most common myths, so that you can talk about this field with real confidence.

• More hype than almost anything
• Conscious or useless?
• Neither

Myth: LLMs understand like we do

A language model predicts the next token, extremely well. Does that count as understanding? People argue about it. For an engineer, the practical point is that it can sound completely sure, and be completely wrong, because sounding right is what it was trained for. So treat its output as a very good draft, not as a fact.

Myth: classic NLP is dead

Plenty of people say that language models made everything else useless. Yet in this course, Naive Bayes took a moment, and a regular expression never made anything up. At high volume, those savings are huge. Your search bar uses BM25. Classic methods remain the fast layer, under the expensive one.

Myth: bigger context solves RAG

When context windows grew, some people said that RAG was over. Just paste everything in. That costs far more, it runs slower, and models do get distracted by irrelevant text. Finding the right few passages still beats sending everything. Retrieval's getting better, rather than going away.

Myth: you must fine-tune

Newcomers often assume that custom means fine tuning. It usually doesn't. Most problems are solved by better prompts, better retrieval, and better examples. Fine tuning changes style and format well, but it doesn't add reliable knowledge. Try those cheaper options first.

Summary