The problem nobody noticed for fifty years
Computers are really good at numbers. Ask a computer to multiply 4,721,903 by 8,112 and you get an answer in a microsecond. Ask a computer to read a sentence like "I can't believe it's not butter" and tell you whether the speaker is surprised, disappointed, or making a joke — and for most of computing history, the answer was: it can't.
Language is messy. It has sarcasm, context, ambiguity, metaphor, and a thousand other things that don't map cleanly onto rules. Natural Language Processing (NLP) is the field of computer science devoted to getting machines to handle that mess.
NLP in one sentence: teaching computers to read, understand, and generate human language.
Why it's hard
Consider a single English sentence:
"The bank is on the river."
A human reads that in a quarter of a second. A computer has to answer:
- Is "bank" the financial institution or the side of a river?
- Which "river" — there are thousands.
- Is "is on" a location, a relationship, something else?
- Is this a statement, a warning, a real-estate ad?
Multiply that by every sentence in every book, website, and tweet ever written, and you see why NLP was stuck for decades.
The stuff NLP tries to do
Before Large Language Models swallowed the field, NLP was a toolbox of individual tasks:
| Task | Example |
|---|---|
| Classification | Is this email spam? Is this review positive? |
| Translation | English → Spanish, Spanish → Japanese |
| Summarization | Boil down a 20-page article into a paragraph |
| Named-entity recognition | "Paris" is a city; "Paris Hilton" is a person |
| Question answering | Given a document, answer a question about it |
| Speech recognition | Turn audio into text |
| Sentiment analysis | Is the author angry, happy, sarcastic? |
Every one of these used to be a separate research specialty with its own algorithms. Modern LLMs do all of them with the same model. That consolidation is the big story of the last five years.
Rules, statistics, and now neural networks
NLP has gone through three big eras:
Rules (1950s–1990s). Humans wrote down the rules of language and programmed them in. "A sentence is a noun phrase followed by a verb phrase..." This fell apart the moment real text showed up — nobody follows the rules.
Statistics (1990s–2010s). Instead of writing rules, count things. If the word "great" appears near "movie" in a million positive reviews, it's probably a positive signal. This worked for simple tasks but couldn't handle meaning or context well.
Neural networks (2010s–now). Let a machine figure out the patterns itself, from enough examples. This is where things exploded. Modern LLMs are the end result of this line of work.
Why this matters for you: every concept you'll learn later — tokens, embeddings, attention, transformers — is an answer to a problem NLP was stuck on. Knowing the problem first makes the solutions stop feeling like magic.
Where NLP sits today
In 2026, "NLP" as a separate field is blurring into "AI" more broadly. One model can read, write, translate, summarize, and reason. The distinctions between "NLP model" and "LLM" and "AI assistant" have mostly collapsed.
But the name sticks around, and the problems are the same. Anytime someone writes "sentiment analysis," "document classification," "text summarization" — that's NLP. It's just that now the answer to almost all of them is: ask an LLM.
What to take away
- NLP is about teaching computers to handle human language.
- Human language is hard because it's ambiguous, contextual, and full of exceptions.
- The field moved from hand-written rules → counting statistics → neural networks.
- Modern LLMs are the latest, most general tool in NLP's toolbox.
Next up: What is Machine Learning? — the idea that lets a computer learn patterns from examples instead of following explicit rules.