Author: Claude AI, under the supervision, prompting and editing by HocTro
Opening
The previous article established that AI is essentially "continuously guessing the next token." That makes a lot of people wonder: that's all? Just guessing the next word? Then how can it solve university-level math, translate Tang Dynasty poetry, write working code, explain philosophy, or summarize a 40-page scientific paper in two minutes?
The answer lies in two things: the internal structure of AI (the neural network and the attention mechanism), and the enormous scale of training. This article explains both — no math required, no technical background needed.
Artificial Neural Networks — Not a Brain, But Inspired by One
A neural network is a mathematical structure built on the idea of how the human brain works — not a simulation of the brain, but inspired by it.
In the human brain, there are billions of neurons. Each neuron connects to thousands of others. When you see a face, a cascade of neurons fires in sequence — one triggering the next, spreading through layer after layer — and eventually your brain recognizes "that's my mother's face." Nobody programmed each step of that. The brain learned how to recognize faces on its own through thousands of exposures.
Artificial neural networks do the same thing with numbers. They consist of many layers of nodes — each node is a small computation unit. A signal (a number) enters the first layer, gets processed, then passes to the next layer, then the next, until the final layer produces an output.
The key detail: every connection between two nodes has a weight — a number that says "how important is this connection?" The entire "intelligence" of AI lives in these billions of weight numbers. When AI learns, what's really happening is that billions of weights are being adjusted until the output becomes more correct.
How Does AI Learn? — Billions of Small Adjustments
Imagine you have a gigantic control panel with a hundred billion dials. Each dial controls one weight in the network. Initially, all the dials are set randomly — the output is meaningless. Feed the system a sentence, ask it to guess the next token, and it's completely wrong.
Then you adjust each dial by a tiny amount — a very small amount — in the direction that makes the output more correct. Repeat that process billions, trillions of times, across hundreds of billions of sentences of text. After enough adjustments, the system starts guessing better — not because anyone programmed in the answers, but because a hundred billion dials have been calibrated through accumulated experience.
This adjustment process is done by an algorithm called backpropagation — it calculates when an output is wrong, which dials are "responsible" for that error, and adjusts those specific dials first. It sounds complex, but the core idea is simple: wrong, then adjust — wrong, then adjust — trillions of times.
A modern AI model like Claude Sonnet has roughly tens of billions of parameters (those "dials"). GPT-4 is estimated to have around 1.8 trillion. Most of AI's intelligence lives in the coordinated interaction of all these numbers — not in any single hand-programmed rule.
The Attention Mechanism — The Real Secret Sauce
This is the part that changed everything. Before 2017, language AI existed but wasn't powerful enough to be truly useful. In 2017, a team of researchers at Google published a paper titled "Attention Is All You Need" — and from that point, language AI entered a new era.
The attention mechanism solves a specific problem: when AI reads a long sentence, it needs to know which word relates to which other word.
Example: "The cat sat on the mat because it was tired."
Who does "it" refer to? The cat or the mat? You know immediately — a mat can't be tired. But how does AI know? The distance between "it" and "cat" is several words — AI has to "look back" far enough and understand that "it" links to "cat," not "mat."
Before the attention mechanism, AI processed sentences left to right, one word at a time — like reading a book with a very short memory — and easily "forgot" things that happened earlier in the sentence. After the attention mechanism, AI can look at the entire sentence at once and compute: which word, for this current token, deserves the most attention?
More precisely: for every token being processed, the attention mechanism computes a score for every other token in the context — that score says "how relevant is this other token to what I'm processing right now?" Highly relevant tokens get more attention in computing the next step; less relevant ones get less.
The result: AI can understand long, complex sentences, maintain meaning across many clauses, know that "it" refers to "cat" and not "mat" — and do this across thousands of words simultaneously, not just within a short sentence.
The architecture built around the attention mechanism is called the Transformer — and this is the foundation of nearly every modern language AI: GPT, Claude, Gemini, LLaMA — all Transformers.
Scale — Why Does Bigger Mean Smarter?
Neural networks and attention mechanisms existed before 2017. But the real explosion in AI capability from 2020 onward came from one more factor: scale.
When you increase the number of parameters (more dials), use more training data, and run more computation — AI becomes smarter in a nonlinear way. It's not that doubling the resources produces double the intelligence — sometimes increasing resources tenfold produces a model that's a hundred times better at specific tasks.
This leads to a surprising phenomenon that researchers call emergent behaviors — capabilities that suddenly appear in large models that nobody programmed, and that couldn't be predicted in advance.
Emergent Behaviors — Skills Nobody Taught
This is the most fascinating — and most mysterious — part.
Imagine a child learning Vietnamese. You teach her vocabulary, grammar, how to structure sentences. At some point — without anyone teaching this directly — the child starts asking questions, then telling stories, then making jokes. The ability to joke was never directly taught; it emerged from having enough command of the language.
AI works the same way. When a model is small, it only does well at what it was explicitly trained for. But when it reaches a sufficient scale, new capabilities suddenly appear — like a phase transition: water at 99°C is still liquid, but at exactly 100°C it suddenly becomes steam. No gradual middle stage — just a sudden shift.
Emergent behaviors that have been documented in large models:
Multi-step reasoning: Small models can't solve problems of the form "if A then B, if B then C, so what does A lead to?" Models above a certain size can suddenly do it, even without being explicitly trained on that type of problem.
In-context learning: You give AI three examples following the same pattern, then present a fourth case — small models are stumped, large models recognize the pattern and answer correctly. No retraining required; just reading the examples within the conversation is enough.
Chain-of-thought reasoning: Large models, when asked to "think step by step before answering," spontaneously generate intermediate reasoning chains and arrive at significantly more accurate answers. Applying the same technique to a small model produces no improvement at all.
Understanding metaphor and humor: Subtle linguistic capabilities — catching sarcasm, recognizing an implied joke, distinguishing literal from figurative meaning — appear in large models and are nearly absent in small ones.
Nobody programmed these skills in. They emerged from the combination of the Transformer architecture, enormous data, and enough parameters.
Genuinely Smart, or Just Sophisticated Statistics?
This question is still open in the AI research community. One camp says: it's all extremely sophisticated statistics — AI doesn't "understand" anything, it just guesses patterns extraordinarily well. The other camp says: when behavior is sufficiently complex and flexible, the line between "statistics" and "intelligence" becomes genuinely unclear.
From a practical standpoint, the question doesn't change much about how you use AI. What matters more is knowing what AI is good at and where it falls short:
- Strong at: synthesizing information, explaining concepts, translation, writing, analysis, coding, answering questions within what it has learned
- Weak at: current events past its training cutoff (without a search tool), independent fact verification, reliably accurate numbers and dates, complex mathematical reasoning without tools
Summing Up
AI isn't intelligent because someone programmed in every answer. It's intelligent because of:
- Artificial neural networks — billions of parameters calibrated through trillions of learning steps
- The attention mechanism — the ability to look at the entire context at once and know what relates to what
- Scale — once big enough, skills nobody explicitly taught begin to emerge on their own
Next article: what does the training process actually look like? Who provides the data, how much does it cost, and how did the technique called RLHF teach AI to be helpful and reasonably polite?
Quick Reference Table
| Concept | Vietnamese Term | Short Definition |
|---|---|---|
| Neural network | Mạng thần kinh nhân tạo | Layers of connected nodes that process numerical signals |
| Weight / parameter | Trọng số / tham số | The adjustable numbers inside every connection |
| Backpropagation | Lan truyền ngược | Algorithm that adjusts weights when output is wrong |
| Attention mechanism | Cơ chế chú ý | Lets AI compute which words relate to which other words |
| Transformer | Kiến trúc Transformer | The architecture behind GPT, Claude, Gemini — uses attention |
| Scale | Quy mô | Parameters + data + computation during training |
| Emergent behavior | Hành vi nổi sinh | A skill that appears in large models without being explicitly taught |
Key Things to Remember
- AI's intelligence isn't programmed answer by answer. It lives in billions of weights, adjusted through trillions of learning steps.
- The attention mechanism was the turning point. It lets AI process the full context at once instead of reading left-to-right and forgetting earlier content.
- Transformer is the foundation of modern AI. GPT, Claude, and Gemini all use this architecture.
- Bigger isn't just faster — it's qualitatively more capable. New skills emerge at scale in ways that can't be predicted in advance.
- Emergent behaviors were never programmed. Multi-step reasoning, in-context learning, understanding humor — all emerged from scale and architecture.
- Next article: How does AI training actually work — the data, the cost, and the human feedback that teaches AI to behave helpfully?