7.2.2 Development History of Large Models

Read This Page as a Map, Not a Memory Test

You do not need to memorize dates. Keep one line in mind:

rules -> statistics -> neural representations -> attention -> scale -> alignment -> tools

Large language models are the result of this long shift, not a sudden invention.

The 15-Stage Big Picture

Stage	What changed	Why it matters to LLMs
1. Turing question	machine intelligence became a concrete question	language became a key test of intelligence
2. Dartmouth AI	AI became a research field	symbolic reasoning dominated early thinking
3. Perceptron	neural learning appeared	first wave of trainable models
4. Expert systems	rules scaled inside narrow domains	showed both value and maintenance pain
5. Backpropagation	multilayer neural nets became trainable	foundation for deep learning
6. LeNet	neural nets worked on real perception tasks	showed representation learning in practice
7. Statistical ML	data-driven methods beat many hand rules	NLP moved toward corpus evidence
8. ImageNet / AlexNet	deep learning won at scale	data + compute + architecture mattered
9. ResNet	very deep networks became trainable	scale became more reliable
10. RNN / LSTM	sequences became neural	language modeling moved beyond n-grams
11. Attention	models could focus on relevant positions	solved part of long-context bottleneck
12. Transformer	attention became the main architecture	parallel training and scaling took off
13. BERT / GPT	pretraining became the shared foundation	one model could transfer to many tasks
14. RLHF / ChatGPT	behavior was aligned with instructions	model capability became product behavior
15. RAG / Agent	models used knowledge and tools	LLMs became application systems

Now zoom in on the language-model line.

Five Language-Model Eras

Era	Core idea	Main limitation
Rule-based systems	humans write language rules	brittle and expensive to maintain
Statistical language models	next word follows observed frequency	sparse data and short context
Neural sequence models	learn vectors and recurrent state	hard to train long dependencies
Transformers	every token can attend to relevant tokens	compute and data cost are high
LLM + alignment	scale pretraining, then tune behavior	hallucination, safety, cost, evaluation

The through line is context. Each era tried to use more context with fewer brittle assumptions.

Lab: Build a Bigram Language Model

This small n-gram model predicts the next word from the current word. It is not powerful, but it shows the statistical idea that came before neural LMs.

from collections import Counter, defaultdict

corpus = [
    "I like learning AI",
    "I like learning Python",
    "You like learning NLP",
    "I like doing projects",
]

next_word_counter = defaultdict(Counter)

for sentence in corpus:
    tokens = sentence.split()
    for current_word, next_word in zip(tokens[:-1], tokens[1:]):
        next_word_counter[current_word][next_word] += 1


def suggest_next(word):
    candidates = next_word_counter[word]
    return candidates.most_common() if candidates else []


print("Common words after I       :", suggest_next("I"))
print("Common words after like    :", suggest_next("like"))
print("Common words after learning:", suggest_next("learning"))

Expected output:

Common words after I       : [('like', 3)]
Common words after like    : [('learning', 3), ('doing', 1)]
Common words after learning: [('AI', 1), ('Python', 1), ('NLP', 1)]

This already feels like autocomplete. But it has three obvious limits:

• it only looks one word back;
• rare combinations have weak statistics;
• it has no semantic representation of the sentence.

Why Neural Models Mattered

Neural language models replaced raw counting with learned representations:

word id -> vector -> context state -> prediction

Word2Vec, GloVe, RNN, LSTM, and GRU made language modeling more flexible. They helped models learn similarity and longer context, but sequential reading still made training slow and long-range memory fragile.

Why Transformer Was the Turning Point

RNNs read mainly step by step. Transformers let tokens directly compare with other tokens through attention:

current token -> attends to relevant tokens -> updated representation

That changed three things:

• training could be more parallel;
• long-range relationships became easier to model;
• scaling parameters, data, and compute became more effective.

This is why BERT, GPT, T5, and later LLMs share the Transformer family tree.

Why Scale Was Not Enough

Large-scale pretraining made models broadly capable, but product behavior still needed another layer:

Need	Technique
follow instructions	instruction tuning
prefer helpful responses	preference learning / RLHF
use current private knowledge	RAG
perform actions	tool calling / Agent loop
reduce unsafe behavior	safety evaluation and guardrails

This is the key modern distinction:

model capability != model behavior

A model can be powerful and still fail to follow policy, cite evidence, or act safely.

What to Remember

Large models belong to NLP history, but they now exceed a narrow NLP boundary. The same architecture and training ideas are used for text, image, speech, code, video, multimodal QA, RAG, and agents.

The practical lesson is:

• rules gave control but poor coverage;
• statistics gave data evidence but short context;
• neural representations gave semantic space;
• Transformer made scale practical;
• alignment, RAG, and tools turned models into systems.

Exercises

1. Add two sentences to the bigram corpus and observe how suggestions change.
2. Why does a bigram model fail on long instructions?
3. Explain why Transformer training is easier to parallelize than RNN training.
4. Give one example where a model has capability but still needs alignment or RAG.
5. Pick one of the 15 stages and explain how it still affects today’s LLM applications.