The evolution of large language models, from ELIZA to AI agents (2026)

When ChatGPT launched in November 2022, it felt to many people as if language AI had appeared overnight. It hadn’t. The systems we use today are the result of more than sixty years of work, with a handful of breakthroughs that changed the pace.

Knowing that history is more than trivia. It explains what these models are good at, why they still make things up, and why the model you choose this year will probably be replaced next year.

What a large language model actually does

A large language model is trained on an enormous amount of text to do one thing: predict the next piece of a sentence. Given “The invoice is due on”, it estimates which words are most likely to follow. Do that billions of times across books, websites, code and documents, and the model picks up grammar, facts, styles of writing and, to a surprising degree, how to reason through a problem.

That design explains both the strengths and the weaknesses. LLMs are very good at anything that looks like language: drafting, summarising, translating, classifying, answering questions and writing code. They are not databases. When a model doesn’t know something, it can still produce a fluent, confident answer that is wrong. That’s what people mean by “hallucination”.

A short history, in six stages

1. Rules (1960s to 1980s)

ELIZA, built at MIT in 1966, imitated a therapist by matching keywords and reflecting them back as questions. It understood nothing, but people still confided in it. Systems of this era followed rules written by hand, and broke as soon as a conversation left the script.

2. Statistics (1990s to 2000s)

Researchers shifted from writing rules to counting. Statistical models learned from real text how likely one word was to follow another, using short sequences called n-grams. This powered spell checkers, early machine translation and predictive text on phones. It worked, but only over a few words of context.

3. Neural networks (2010s)

Neural networks let models learn richer patterns. In 2013, Word2Vec represented words as points in space, so that words with similar meanings sat close together. Recurrent neural networks could read a sentence word by word and carry some memory forward. They were better, but slow to train and forgetful over long passages.

4. The Transformer (2017)

In 2017 a team at Google published Attention Is All You Need, which introduced the Transformer. Instead of reading word by word, a Transformer looks at a whole passage at once and learns which words matter to each other. It also trains efficiently on modern chips. Almost every major language model since is built on it.

In 2018 Google released BERT, which was good at understanding text and quickly improved search, and OpenAI released the first GPT, which was good at generating it. GPT-3 in 2020, with 175 billion parameters, showed that simply making these models bigger and training them on more text produced abilities nobody had explicitly programmed.

5. Chat goes mainstream (2022 to 2023)

ChatGPT, released in November 2022, wrapped a GPT model in a simple chat window and was trained to follow instructions and hold a conversation. It reached a mass audience faster than almost any product before it; by October 2025 OpenAI said it had 800 million weekly users. GPT-4 followed in March 2023, alongside Anthropic’s Claude, Google’s Bard (later Gemini) and Meta’s openly released Llama models.

6. From answers to actions (2024 to today)

Since 2024 the race has been less about size and more about usefulness:

• Multimodal models read and produce images, audio and documents as well as text. You can show a model a photo of a damaged part or a scanned contract.
• Reasoning models, starting with OpenAI’s o1 in 2024, work through a problem step by step before answering. They are much stronger at maths, analysis and code.
• Retrieval-augmented generation (RAG) connects a model to your own documents, so answers come from current, approved sources instead of memory. It’s the main way companies reduce made-up answers. We explain it in our guide to building a RAG knowledge base.
• Open-weight models, such as Meta’s Llama family and DeepSeek’s releases in early 2025, showed that capable models could be run privately and far more cheaply than many assumed.
• Agents and tool use. Models can now call other software: search the web, query a database, update a CRM, fill in a form. Shared standards such as the Model Context Protocol, introduced in late 2024, make it easier to connect them to business systems.

The problems that haven’t gone away

• Made-up answers. Better models hallucinate less, but none are immune. Grounding answers in your own sources and checking anything that matters is still essential.
• Bias. Models learn from human writing and can reflect its biases, which matters when they help screen candidates, assess customers or write about people.
• Data and privacy. Anything you send to a model goes somewhere. Companies need clear rules on which tools staff may use and what data can go into them.
• Cost at scale. A single question is cheap. Millions of them, or agents that run for minutes at a time, are not. Choosing the right size of model for each task matters.
• Constant change. Providers release new models every few months and retire old ones. Systems built around one specific model need testing and updating whenever it changes.

What this means for a business

The history points to a practical rule: build around the job, not the model. Decide what the task is, what a good result looks like and how you’ll measure it. Then test two or three models against your own examples, put the winner into the tools your team already uses, and expect to swap it out within a year or two.

Most of the value companies get from LLMs today comes from ordinary, repeated work: answering customer questions from a knowledge base, drafting proposals, summarising calls, pulling data out of documents. Our guide to choosing an LLM for your business covers how to compare models fairly. If you’re deciding where they would pay off first, that’s the kind of question we work through in AI implementation and automation.