Not a new branch
The shortest true version of this history: each era of AI grew out of the one before it rather than replacing it. An LLM is a deep neural network; a deep neural network is machine learning; machine learning is AI. The words are not competing options a vendor picks between. They are nested, and the newest one sits inside all the older ones.
So the useful picture is a tree. The roots are the older fields AI borrowed from, none of which were about computers being clever. The trunk is the sequence of eras, each one a ring that the next one grew around. The branches are the named models you can use this afternoon.
Reading it this way pays off immediately. When the next capability is announced, the question to ask is which ring it grew from - and the answer is almost always an older ring at a new scale.
The roots: seven other fields
AI was assembled out of work that was already underway elsewhere. Seven fields contributed the pieces:
- Mathematics - logic, probability, and optimisation: the machinery for representing a claim and for improving a guess.
- Computer science - algorithms, data structures, and the machines to run them on.
- Statistics - models, inference, and experimental design: how to learn from a sample and how to know whether you did.
- Cognitive science - what perception, memory, and reasoning look like as processes.
- Linguistics - grammar, meaning, and the structure inside a sentence.
- Neuroscience - the brain as a network of simple units that learn by changing their connections.
- Psychology - behaviour, learning, and how decisions are actually made.
These are not historical trivia. Each root is still load-bearing in a model you used this week: probability is what makes a model choose the next word, linguistics is why text gets split into tokens before anything else happens, statistics is the whole training procedure, and optimisation is the part that does the learning.
One root deserves a caveat. The word neural is an analogy, not an engineering claim. An artificial neuron adds up some numbers, multiplies them by weights, and passes the result on. That is a useful borrowing of a metaphor from neuroscience, and it is nothing like a brain cell.
1950s to 1970s: hand-written rules
The first working theory of machine intelligence was that thinking is logic plus search. A person writes down the rules; the machine applies them faithfully and explores the possibilities faster than a person can. This approach has a name: symbolic AI, because it manipulates symbols and rules rather than numbers learned from data.
The showpiece was the expert system - a program holding a few hundred IF-THEN rules collected from specialists in a narrow field. A 1970s medical version asked a doctor a series of questions and worked down its rule set to a suggested diagnosis. Within its narrow subject it performed respectably, and every answer could be traced back to the rules that produced it.
Then the approach hit a wall that has nothing to do with computing power. Somebody has to write the rules, and for the interesting problems nobody can. Write the rules that recognise a friend's face in a photograph. Write the rules that translate a paragraph of English into Japanese. People do both without being able to say how, which means there is no list to hand over. Funding and interest collapsed twice over that gap, in periods still called the AI winters.
The ring is still there, though. Search over possible moves, planners, and rule systems run inside plenty of modern software, including the part of an agent that decides which tool to reach for. The era did not fail. It ran out of problems it could express.
1980s to 2000s: learn it from data
The next ring inverted the job. Instead of writing the rules, write a program that derives the rules from examples. That is machine learning: the model is fitted to data rather than authored by hand.
A spam filter is the clean example. The rule-writing version needs a human to list the phrases that mean spam, and it goes stale the week spammers change wording. The learning version gets ten thousand emails already marked spam or not spam, works out which words move the odds, and can be refitted on a newer pile next month without anyone editing a rule.
This era produced the toolbox that still runs most of the world's practical prediction: decision trees (a chain of yes/no splits), support vector machines (find the boundary that separates two groups with the widest margin), and ensemble methods (combine many mediocre models into one good one, because their mistakes cancel out).
It also had one expensive weakness: feature engineering. A person had to decide, in advance, what to measure. For email, word counts are an obvious answer. For a photograph of a cat, deciding what to measure is the problem, and no committee was going to solve it. That unfinished job is exactly what the next ring took on.
Rules written by hand, then rules learned from data, then the features learned too. Each ring took over a job the previous one had left to a human.- the pattern the whole trunk follows
Neural networks: the idea that had to wait
A neural network is a stack of layers of those simple weighted-sum units. What makes it different from the toolbox above is that it learns the measurements as well as the answer. Show it enough photographs and the early layers settle on edges, the middle layers on shapes and textures, and the later layers on whole faces. Nobody wrote down "look for edges first". The layers arranged themselves that way because it made the predictions better.
None of this was new in the 2010s. Working neural networks existed in the 1960s, the training method that makes many-layered ones practical was established in the 1980s, and the design that finally let a network carry information across a long sentence arrived in 1997. For most of those decades neural networks were a minority position that lost bake-offs to the simpler methods.
Three things kept them down, and none of them was the idea itself: there was not enough labelled data, there was not enough compute to train anything deep, and deep stacks were unstable to train even when you had both. The idea was right and early, which in practice is indistinguishable from wrong until the surroundings catch up.
The 2010s: three things arrive at once
The surroundings caught up around 2012, when a neural network won a large image-recognition contest by a margin that ended the argument rather than continuing it. Three separate supplies had arrived together:
- Data - the web, plus large hand-labelled image and text collections. Millions of examples instead of thousands.
- Compute - graphics chips, built to do the same arithmetic on thousands of numbers at once, which is precisely the shape of neural network training.
- Training know-how - a decade of unglamorous fixes to how networks are initialised, activated, and regularised, which together made deep stacks trainable rather than temperamental.
That combination is what deep learning names: neural networks with many layers, trained at a scale that was previously out of reach. "Deep" is literally a count of layers, not a claim about depth of thought. Vision went to convolutional networks; sequences - text, speech - went to recurrent networks and their longer-memory variants.
Those sequence models also set up the next problem. They read one step at a time, in order, because step five depended on step four. That made training slow to the point of capping how much text anyone could use, and it made the link between the first word and the five-hundredth faint by the time it arrived.
2017: attention changes everything
The transformer, published in 2017, replaced reading in order with a mechanism called attention: every position in the text looks directly at every other position and weights how much each one matters for the word being computed. In "the trophy did not fit in the suitcase because it was too big", the word "it" can look straight at "trophy" instead of remembering it through a chain of intervening steps.
Two consequences followed, and the second one is the reason 2017 is the hinge of this whole story:
- Distance stopped mattering. Word one and word five hundred are a single hop apart, so a long document holds together instead of fading.
- Training became parallel. With no strict left-to-right dependency during training, the whole sequence goes through the hardware at once - which turned "more text" from a scheduling impossibility into a budget line.
Once that was true, a pattern showed up that mattered even more than the architecture: making the model bigger, the dataset bigger, and the compute budget bigger produced predictably better results. Predictability is what changed the industry. A capability you can forecast from a spending plan gets funded, and the next few years were largely that forecast being spent.
Worth noting for later: nothing about the transformer is specific to language. The same design now sits under image, audio, video, and code models, which is why today's models can handle several of those at once.
2018 to 2020: pretrain once, adapt often
Until this point, a model was built for a task. Sentiment classifier, translator, spam filter - each one trained from scratch on its own labelled dataset, and each one useless outside its job.
The new pattern split that in two. First pretrain one large model on an enormous amount of ordinary text, on the self-labelling task of predicting the next word: the correct answer is simply the word that came next, so no human has to annotate anything. Then adapt the finished model to a specific job, either by fine-tuning it on a small dataset or, increasingly, by just describing the job in the prompt. How a model learns follows that lifecycle stage by stage.
A model built to be adapted like this is a foundation model: one expensive artifact, many downstream uses. By 2020 the second half of the pattern had become the surprising part - a large enough pretrained model could handle a task it had never been fine-tuned for, given a description and a couple of examples in the prompt. That is the moment AI turned into something you type into rather than something you train.
The same years produced the other half of today's vocabulary. Models that compose new content - text, images, audio, code - rather than sorting existing content into buckets are what Gen AI names. Generation was not a new goal; it became a practical one when pretraining at scale made the output good enough to keep.
The branches today
Put the trunk back together and an LLM is fully described by its lineage: a transformer, trained at deep-learning scale, on the machine-learning premise that rules come from data, inside the project that started with logic and search in the 1950s. The "large" is doing real work in the name - it refers both to the number of weights and to the volume of text they were fitted to.
The branches are the products: ChatGPT and the GPT models, Claude, Gemini, Llama, Mistral, and more arriving each quarter. Different labs, different training choices, different licences, one trunk. Inside an LLM traces a single question through one of them.
What is growing now is the same pattern continuing rather than a break from it:
- Several kinds of input in one model - text, images, audio, and video handled by one set of weights, because attention never cared what the tokens represented.
- Longer context - more of your material in front of the model at once, which is an engineering fight about cost, not a new idea.
- Tools and agents - the model asks for an action, your code runs it, the result comes back. The oldest ring, search over possible next moves, wrapped around the newest one.
- Smaller models catching up - this year's cheap model matching last year's expensive one, which is what makes any of this affordable to run.
LLMs are not a new branch of AI. They are the latest bloom on a long-growing tree.- the one sentence worth keeping
Every era in one line
The vocabulary in order, so the names stop competing with each other:
- Symbolic AI - intelligence as rules and logic, written by people.
- Expert system - a 1970s program that applied a few hundred hand-collected IF-THEN rules in one narrow field.
- AI winter - either of the two periods when the gap between promises and results cut off funding and attention.
- Machine learning - the rules are derived from examples instead of written down.
- Feature engineering - a person deciding in advance what to measure about each example. The job deep learning removed.
- Neural network - layers of simple weighted-sum units that learn the measurements as well as the answer.
- Deep learning - neural networks with many layers, trained on large data with parallel hardware.
- Transformer - the 2017 design whose attention mechanism lets every position read every other position, and whose training runs in parallel.
- Foundation model - one large pretrained model built to be adapted to many tasks rather than trained for one.
- Generative AI - models that compose new content rather than classify existing content.
- Large language model - a foundation model for text: a transformer trained on a very large body of text to continue it.